diff --git a/README.md b/README.md index a94a212ad..ba1e844b3 100644 --- a/README.md +++ b/README.md @@ -142,7 +142,7 @@ You can install ComfyUI in Apple Mac silicon (M1 or M2) with any recent macOS ve 1. Install pytorch nightly. For instructions, read the [Accelerated PyTorch training on Mac](https://developer.apple.com/metal/pytorch/) Apple Developer guide (make sure to install the latest pytorch nightly). 1. Follow the [ComfyUI manual installation](#manual-install-windows-linux) instructions for Windows and Linux. 1. Install the ComfyUI [dependencies](#dependencies). If you have another Stable Diffusion UI [you might be able to reuse the dependencies](#i-already-have-another-ui-for-stable-diffusion-installed-do-i-really-have-to-install-all-of-these-dependencies). -1. Launch ComfyUI by running `python main.py --force-fp16`. Note that --force-fp16 will only work if you installed the latest pytorch nightly. +1. Launch ComfyUI by running `python main.py` > **Note**: Remember to add your models, VAE, LoRAs etc. to the corresponding Comfy folders, as discussed in [ComfyUI manual installation](#manual-install-windows-linux). diff --git a/comfy/cli_args.py b/comfy/cli_args.py index 757fc245f..353bb51e7 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -129,7 +129,7 @@ if args.disable_auto_launch: args.auto_launch = False import logging -logging_level = logging.WARNING +logging_level = logging.INFO if args.verbose: logging_level = logging.DEBUG diff --git a/comfy/controlnet.py b/comfy/controlnet.py index a5e7b23f7..b6941d8c4 100644 --- a/comfy/controlnet.py +++ b/comfy/controlnet.py @@ -201,7 +201,7 @@ class ControlNet(ControlBase): super().cleanup() class ControlLoraOps: - class Linear(torch.nn.Module): + class Linear(torch.nn.Module, comfy.ops.CastWeightBiasOp): def __init__(self, in_features: int, out_features: int, bias: bool = True, device=None, dtype=None) -> None: factory_kwargs = {'device': device, 'dtype': dtype} @@ -220,7 +220,7 @@ class ControlLoraOps: else: return torch.nn.functional.linear(input, weight, bias) - class Conv2d(torch.nn.Module): + class Conv2d(torch.nn.Module, comfy.ops.CastWeightBiasOp): def __init__( self, in_channels, @@ -432,7 +432,7 @@ def load_controlnet(ckpt_path, model=None): logging.warning("missing controlnet keys: {}".format(missing)) if len(unexpected) > 0: - logging.info("unexpected controlnet keys: {}".format(unexpected)) + logging.debug("unexpected controlnet keys: {}".format(unexpected)) global_average_pooling = False filename = os.path.splitext(ckpt_path)[0] @@ -545,6 +545,6 @@ def load_t2i_adapter(t2i_data): logging.warning("t2i missing {}".format(missing)) if len(unexpected) > 0: - logging.info("t2i unexpected {}".format(unexpected)) + logging.debug("t2i unexpected {}".format(unexpected)) return T2IAdapter(model_ad, model_ad.input_channels, compression_ratio, upscale_algorithm) diff --git a/comfy/diffusers_convert.py b/comfy/diffusers_convert.py index 18398cb32..ed2a45fea 100644 --- a/comfy/diffusers_convert.py +++ b/comfy/diffusers_convert.py @@ -178,7 +178,7 @@ def convert_vae_state_dict(vae_state_dict): for k, v in new_state_dict.items(): for weight_name in weights_to_convert: if f"mid.attn_1.{weight_name}.weight" in k: - logging.info(f"Reshaping {k} for SD format") + logging.debug(f"Reshaping {k} for SD format") new_state_dict[k] = reshape_weight_for_sd(v) return new_state_dict @@ -206,6 +206,21 @@ textenc_pattern = re.compile("|".join(protected.keys())) # Ordering is from https://github.com/pytorch/pytorch/blob/master/test/cpp/api/modules.cpp code2idx = {"q": 0, "k": 1, "v": 2} +# This function exists because at the time of writing torch.cat can't do fp8 with cuda +def cat_tensors(tensors): + x = 0 + for t in tensors: + x += t.shape[0] + + shape = [x] + list(tensors[0].shape)[1:] + out = torch.empty(shape, device=tensors[0].device, dtype=tensors[0].dtype) + + x = 0 + for t in tensors: + out[x:x + t.shape[0]] = t + x += t.shape[0] + + return out def convert_text_enc_state_dict_v20(text_enc_dict, prefix=""): new_state_dict = {} @@ -249,13 +264,13 @@ def convert_text_enc_state_dict_v20(text_enc_dict, prefix=""): if None in tensors: raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing") relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre) - new_state_dict[relabelled_key + ".in_proj_weight"] = torch.cat(tensors) + new_state_dict[relabelled_key + ".in_proj_weight"] = cat_tensors(tensors) for k_pre, tensors in capture_qkv_bias.items(): if None in tensors: raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing") relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre) - new_state_dict[relabelled_key + ".in_proj_bias"] = torch.cat(tensors) + new_state_dict[relabelled_key + ".in_proj_bias"] = cat_tensors(tensors) return new_state_dict diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.py index 761c2e0ef..7af016829 100644 --- a/comfy/k_diffusion/sampling.py +++ b/comfy/k_diffusion/sampling.py @@ -748,7 +748,7 @@ def sample_lcm(model, x, sigmas, extra_args=None, callback=None, disable=None, n x = denoised if sigmas[i + 1] > 0: - x += sigmas[i + 1] * noise_sampler(sigmas[i], sigmas[i + 1]) + x = model.inner_model.inner_model.model_sampling.noise_scaling(sigmas[i + 1], noise_sampler(sigmas[i], sigmas[i + 1]), x) return x diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index 674364e72..4ca466d9a 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -95,7 +95,7 @@ class SC_Prior(LatentFormat): class SC_B(LatentFormat): def __init__(self): - self.scale_factor = 1.0 + self.scale_factor = 1.0 / 0.43 self.latent_rgb_factors = [ [ 0.1121, 0.2006, 0.1023], [-0.2093, -0.0222, -0.0195], diff --git a/comfy/ldm/cascade/stage_a.py b/comfy/ldm/cascade/stage_a.py index 260ccfc0b..ca8867eaf 100644 --- a/comfy/ldm/cascade/stage_a.py +++ b/comfy/ldm/cascade/stage_a.py @@ -163,11 +163,9 @@ class ResBlock(nn.Module): class StageA(nn.Module): - def __init__(self, levels=2, bottleneck_blocks=12, c_hidden=384, c_latent=4, codebook_size=8192, - scale_factor=0.43): # 0.3764 + def __init__(self, levels=2, bottleneck_blocks=12, c_hidden=384, c_latent=4, codebook_size=8192): super().__init__() self.c_latent = c_latent - self.scale_factor = scale_factor c_levels = [c_hidden // (2 ** i) for i in reversed(range(levels))] # Encoder blocks @@ -214,12 +212,11 @@ class StageA(nn.Module): x = self.down_blocks(x) if quantize: qe, (vq_loss, commit_loss), indices = self.vquantizer.forward(x, dim=1) - return qe / self.scale_factor, x / self.scale_factor, indices, vq_loss + commit_loss * 0.25 + return qe, x, indices, vq_loss + commit_loss * 0.25 else: - return x / self.scale_factor + return x def decode(self, x): - x = x * self.scale_factor x = self.up_blocks(x) x = self.out_block(x) return x diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.py index 48399bc07..f116efee3 100644 --- a/comfy/ldm/modules/attention.py +++ b/comfy/ldm/modules/attention.py @@ -4,6 +4,7 @@ import torch.nn.functional as F from torch import nn, einsum from einops import rearrange, repeat from typing import Optional, Any +import logging from .diffusionmodules.util import checkpoint, AlphaBlender, timestep_embedding from .sub_quadratic_attention import efficient_dot_product_attention @@ -20,7 +21,7 @@ ops = comfy.ops.disable_weight_init # CrossAttn precision handling if args.dont_upcast_attention: - print("disabling upcasting of attention") + logging.info("disabling upcasting of attention") _ATTN_PRECISION = "fp16" else: _ATTN_PRECISION = "fp32" @@ -274,12 +275,12 @@ def attention_split(q, k, v, heads, mask=None): model_management.soft_empty_cache(True) if cleared_cache == False: cleared_cache = True - print("out of memory error, emptying cache and trying again") + logging.warning("out of memory error, emptying cache and trying again") continue steps *= 2 if steps > 64: raise e - print("out of memory error, increasing steps and trying again", steps) + logging.warning("out of memory error, increasing steps and trying again {}".format(steps)) else: raise e @@ -351,17 +352,17 @@ def attention_pytorch(q, k, v, heads, mask=None): optimized_attention = attention_basic if model_management.xformers_enabled(): - print("Using xformers cross attention") + logging.info("Using xformers cross attention") optimized_attention = attention_xformers elif model_management.pytorch_attention_enabled(): - print("Using pytorch cross attention") + logging.info("Using pytorch cross attention") optimized_attention = attention_pytorch else: if args.use_split_cross_attention: - print("Using split optimization for cross attention") + logging.info("Using split optimization for cross attention") optimized_attention = attention_split else: - print("Using sub quadratic optimization for cross attention, if you have memory or speed issues try using: --use-split-cross-attention") + logging.info("Using sub quadratic optimization for cross attention, if you have memory or speed issues try using: --use-split-cross-attention") optimized_attention = attention_sub_quad optimized_attention_masked = optimized_attention diff --git a/comfy/ldm/modules/diffusionmodules/model.py b/comfy/ldm/modules/diffusionmodules/model.py index cc81c1f23..fabc5c5e5 100644 --- a/comfy/ldm/modules/diffusionmodules/model.py +++ b/comfy/ldm/modules/diffusionmodules/model.py @@ -5,6 +5,7 @@ import torch.nn as nn import numpy as np from einops import rearrange from typing import Optional, Any +import logging from comfy import model_management import comfy.ops @@ -190,7 +191,7 @@ def slice_attention(q, k, v): steps *= 2 if steps > 128: raise e - print("out of memory error, increasing steps and trying again", steps) + logging.warning("out of memory error, increasing steps and trying again {}".format(steps)) return r1 @@ -235,7 +236,7 @@ def pytorch_attention(q, k, v): out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False) out = out.transpose(2, 3).reshape(B, C, H, W) except model_management.OOM_EXCEPTION as e: - print("scaled_dot_product_attention OOMed: switched to slice attention") + logging.warning("scaled_dot_product_attention OOMed: switched to slice attention") out = slice_attention(q.view(B, -1, C), k.view(B, -1, C).transpose(1, 2), v.view(B, -1, C).transpose(1, 2)).reshape(B, C, H, W) return out @@ -268,13 +269,13 @@ class AttnBlock(nn.Module): padding=0) if model_management.xformers_enabled_vae(): - print("Using xformers attention in VAE") + logging.info("Using xformers attention in VAE") self.optimized_attention = xformers_attention elif model_management.pytorch_attention_enabled(): - print("Using pytorch attention in VAE") + logging.info("Using pytorch attention in VAE") self.optimized_attention = pytorch_attention else: - print("Using split attention in VAE") + logging.info("Using split attention in VAE") self.optimized_attention = normal_attention def forward(self, x): @@ -562,7 +563,7 @@ class Decoder(nn.Module): block_in = ch*ch_mult[self.num_resolutions-1] curr_res = resolution // 2**(self.num_resolutions-1) self.z_shape = (1,z_channels,curr_res,curr_res) - print("Working with z of shape {} = {} dimensions.".format( + logging.debug("Working with z of shape {} = {} dimensions.".format( self.z_shape, np.prod(self.z_shape))) # z to block_in diff --git a/comfy/ldm/modules/diffusionmodules/openaimodel.py b/comfy/ldm/modules/diffusionmodules/openaimodel.py index cf89ae017..d782eff31 100644 --- a/comfy/ldm/modules/diffusionmodules/openaimodel.py +++ b/comfy/ldm/modules/diffusionmodules/openaimodel.py @@ -4,6 +4,7 @@ import torch as th import torch.nn as nn import torch.nn.functional as F from einops import rearrange +import logging from .util import ( checkpoint, @@ -359,7 +360,7 @@ def apply_control(h, control, name): try: h += ctrl except: - print("warning control could not be applied", h.shape, ctrl.shape) + logging.warning("warning control could not be applied {} {}".format(h.shape, ctrl.shape)) return h class UNetModel(nn.Module): @@ -496,7 +497,7 @@ class UNetModel(nn.Module): if isinstance(self.num_classes, int): self.label_emb = nn.Embedding(num_classes, time_embed_dim, dtype=self.dtype, device=device) elif self.num_classes == "continuous": - print("setting up linear c_adm embedding layer") + logging.debug("setting up linear c_adm embedding layer") self.label_emb = nn.Linear(1, time_embed_dim) elif self.num_classes == "sequential": assert adm_in_channels is not None diff --git a/comfy/ldm/modules/sub_quadratic_attention.py b/comfy/ldm/modules/sub_quadratic_attention.py index cb0896b0d..1bc4138c3 100644 --- a/comfy/ldm/modules/sub_quadratic_attention.py +++ b/comfy/ldm/modules/sub_quadratic_attention.py @@ -14,6 +14,7 @@ import torch from torch import Tensor from torch.utils.checkpoint import checkpoint import math +import logging try: from typing import Optional, NamedTuple, List, Protocol @@ -170,7 +171,7 @@ def _get_attention_scores_no_kv_chunking( attn_probs = attn_scores.softmax(dim=-1) del attn_scores except model_management.OOM_EXCEPTION: - print("ran out of memory while running softmax in _get_attention_scores_no_kv_chunking, trying slower in place softmax instead") + logging.warning("ran out of memory while running softmax in _get_attention_scores_no_kv_chunking, trying slower in place softmax instead") attn_scores -= attn_scores.max(dim=-1, keepdim=True).values torch.exp(attn_scores, out=attn_scores) summed = torch.sum(attn_scores, dim=-1, keepdim=True) diff --git a/comfy/lora.py b/comfy/lora.py index 637380d54..096285bba 100644 --- a/comfy/lora.py +++ b/comfy/lora.py @@ -21,6 +21,12 @@ def load_lora(lora, to_load): alpha = lora[alpha_name].item() loaded_keys.add(alpha_name) + dora_scale_name = "{}.dora_scale".format(x) + dora_scale = None + if dora_scale_name in lora.keys(): + dora_scale = lora[dora_scale_name] + loaded_keys.add(dora_scale_name) + regular_lora = "{}.lora_up.weight".format(x) diffusers_lora = "{}_lora.up.weight".format(x) transformers_lora = "{}.lora_linear_layer.up.weight".format(x) @@ -44,7 +50,7 @@ def load_lora(lora, to_load): if mid_name is not None and mid_name in lora.keys(): mid = lora[mid_name] loaded_keys.add(mid_name) - patch_dict[to_load[x]] = ("lora", (lora[A_name], lora[B_name], alpha, mid)) + patch_dict[to_load[x]] = ("lora", (lora[A_name], lora[B_name], alpha, mid, dora_scale)) loaded_keys.add(A_name) loaded_keys.add(B_name) @@ -65,7 +71,7 @@ def load_lora(lora, to_load): loaded_keys.add(hada_t1_name) loaded_keys.add(hada_t2_name) - patch_dict[to_load[x]] = ("loha", (lora[hada_w1_a_name], lora[hada_w1_b_name], alpha, lora[hada_w2_a_name], lora[hada_w2_b_name], hada_t1, hada_t2)) + patch_dict[to_load[x]] = ("loha", (lora[hada_w1_a_name], lora[hada_w1_b_name], alpha, lora[hada_w2_a_name], lora[hada_w2_b_name], hada_t1, hada_t2, dora_scale)) loaded_keys.add(hada_w1_a_name) loaded_keys.add(hada_w1_b_name) loaded_keys.add(hada_w2_a_name) @@ -117,7 +123,7 @@ def load_lora(lora, to_load): loaded_keys.add(lokr_t2_name) if (lokr_w1 is not None) or (lokr_w2 is not None) or (lokr_w1_a is not None) or (lokr_w2_a is not None): - patch_dict[to_load[x]] = ("lokr", (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2)) + patch_dict[to_load[x]] = ("lokr", (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2, dora_scale)) #glora a1_name = "{}.a1.weight".format(x) @@ -125,7 +131,7 @@ def load_lora(lora, to_load): b1_name = "{}.b1.weight".format(x) b2_name = "{}.b2.weight".format(x) if a1_name in lora: - patch_dict[to_load[x]] = ("glora", (lora[a1_name], lora[a2_name], lora[b1_name], lora[b2_name], alpha)) + patch_dict[to_load[x]] = ("glora", (lora[a1_name], lora[a2_name], lora[b1_name], lora[b2_name], alpha, dora_scale)) loaded_keys.add(a1_name) loaded_keys.add(a2_name) loaded_keys.add(b1_name) diff --git a/comfy/model_base.py b/comfy/model_base.py index a2514ca5e..898adb665 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -66,9 +66,10 @@ class BaseModel(torch.nn.Module): self.adm_channels = unet_config.get("adm_in_channels", None) if self.adm_channels is None: self.adm_channels = 0 - self.inpaint_model = False - logging.warning("model_type {}".format(model_type.name)) - logging.info("adm {}".format(self.adm_channels)) + + self.concat_keys = () + logging.info("model_type {}".format(model_type.name)) + logging.debug("adm {}".format(self.adm_channels)) def apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs): sigma = t @@ -107,8 +108,7 @@ class BaseModel(torch.nn.Module): def extra_conds(self, **kwargs): out = {} - if self.inpaint_model: - concat_keys = ("mask", "masked_image") + if len(self.concat_keys) > 0: cond_concat = [] denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None)) concat_latent_image = kwargs.get("concat_latent_image", None) @@ -125,24 +125,16 @@ class BaseModel(torch.nn.Module): concat_latent_image = utils.resize_to_batch_size(concat_latent_image, noise.shape[0]) - if len(denoise_mask.shape) == len(noise.shape): - denoise_mask = denoise_mask[:,:1] + if denoise_mask is not None: + if len(denoise_mask.shape) == len(noise.shape): + denoise_mask = denoise_mask[:,:1] - denoise_mask = denoise_mask.reshape((-1, 1, denoise_mask.shape[-2], denoise_mask.shape[-1])) - if denoise_mask.shape[-2:] != noise.shape[-2:]: - denoise_mask = utils.common_upscale(denoise_mask, noise.shape[-1], noise.shape[-2], "bilinear", "center") - denoise_mask = utils.resize_to_batch_size(denoise_mask.round(), noise.shape[0]) + denoise_mask = denoise_mask.reshape((-1, 1, denoise_mask.shape[-2], denoise_mask.shape[-1])) + if denoise_mask.shape[-2:] != noise.shape[-2:]: + denoise_mask = utils.common_upscale(denoise_mask, noise.shape[-1], noise.shape[-2], "bilinear", "center") + denoise_mask = utils.resize_to_batch_size(denoise_mask.round(), noise.shape[0]) - def blank_inpaint_image_like(latent_image): - blank_image = torch.ones_like(latent_image) - # these are the values for "zero" in pixel space translated to latent space - blank_image[:,0] *= 0.8223 - blank_image[:,1] *= -0.6876 - blank_image[:,2] *= 0.6364 - blank_image[:,3] *= 0.1380 - return blank_image - - for ck in concat_keys: + for ck in self.concat_keys: if denoise_mask is not None: if ck == "mask": cond_concat.append(denoise_mask.to(device)) @@ -152,7 +144,7 @@ class BaseModel(torch.nn.Module): if ck == "mask": cond_concat.append(torch.ones_like(noise)[:,:1]) elif ck == "masked_image": - cond_concat.append(blank_inpaint_image_like(noise)) + cond_concat.append(self.blank_inpaint_image_like(noise)) data = torch.cat(cond_concat, dim=1) out['c_concat'] = comfy.conds.CONDNoiseShape(data) @@ -221,7 +213,16 @@ class BaseModel(torch.nn.Module): return unet_state_dict def set_inpaint(self): - self.inpaint_model = True + self.concat_keys = ("mask", "masked_image") + def blank_inpaint_image_like(latent_image): + blank_image = torch.ones_like(latent_image) + # these are the values for "zero" in pixel space translated to latent space + blank_image[:,0] *= 0.8223 + blank_image[:,1] *= -0.6876 + blank_image[:,2] *= 0.6364 + blank_image[:,3] *= 0.1380 + return blank_image + self.blank_inpaint_image_like = blank_inpaint_image_like def memory_required(self, input_shape): if comfy.model_management.xformers_enabled() or comfy.model_management.pytorch_attention_flash_attention(): @@ -380,6 +381,36 @@ class SVD_img2vid(BaseModel): out['num_video_frames'] = comfy.conds.CONDConstant(noise.shape[0]) return out +class SV3D_u(SVD_img2vid): + def encode_adm(self, **kwargs): + augmentation = kwargs.get("augmentation_level", 0) + + out = [] + out.append(self.embedder(torch.flatten(torch.Tensor([augmentation])))) + + flat = torch.flatten(torch.cat(out)).unsqueeze(dim=0) + return flat + +class SV3D_p(SVD_img2vid): + def __init__(self, model_config, model_type=ModelType.V_PREDICTION_EDM, device=None): + super().__init__(model_config, model_type, device=device) + self.embedder_512 = Timestep(512) + + def encode_adm(self, **kwargs): + augmentation = kwargs.get("augmentation_level", 0) + elevation = kwargs.get("elevation", 0) #elevation and azimuth are in degrees here + azimuth = kwargs.get("azimuth", 0) + noise = kwargs.get("noise", None) + + out = [] + out.append(self.embedder(torch.flatten(torch.Tensor([augmentation])))) + out.append(self.embedder_512(torch.deg2rad(torch.fmod(torch.flatten(90 - torch.Tensor([elevation])), 360.0)))) + out.append(self.embedder_512(torch.deg2rad(torch.fmod(torch.flatten(torch.Tensor([azimuth])), 360.0)))) + + out = list(map(lambda a: utils.resize_to_batch_size(a, noise.shape[0]), out)) + return torch.cat(out, dim=1) + + class Stable_Zero123(BaseModel): def __init__(self, model_config, model_type=ModelType.EPS, device=None, cc_projection_weight=None, cc_projection_bias=None): super().__init__(model_config, model_type, device=device) @@ -442,6 +473,40 @@ class SD_X4Upscaler(BaseModel): out['y'] = comfy.conds.CONDRegular(noise_level) return out +class IP2P: + def extra_conds(self, **kwargs): + out = {} + + image = kwargs.get("concat_latent_image", None) + noise = kwargs.get("noise", None) + device = kwargs["device"] + + if image is None: + image = torch.zeros_like(noise) + + if image.shape[1:] != noise.shape[1:]: + image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center") + + image = utils.resize_to_batch_size(image, noise.shape[0]) + + out['c_concat'] = comfy.conds.CONDNoiseShape(self.process_ip2p_image_in(image)) + adm = self.encode_adm(**kwargs) + if adm is not None: + out['y'] = comfy.conds.CONDRegular(adm) + return out + +class SD15_instructpix2pix(IP2P, BaseModel): + def __init__(self, model_config, model_type=ModelType.EPS, device=None): + super().__init__(model_config, model_type, device=device) + self.process_ip2p_image_in = lambda image: image + +class SDXL_instructpix2pix(IP2P, SDXL): + def __init__(self, model_config, model_type=ModelType.EPS, device=None): + super().__init__(model_config, model_type, device=device) + # self.process_ip2p_image_in = lambda image: comfy.latent_formats.SDXL().process_in(image) + self.process_ip2p_image_in = lambda image: image + + class StableCascade_C(BaseModel): def __init__(self, model_config, model_type=ModelType.STABLE_CASCADE, device=None): super().__init__(model_config, model_type, device=device, unet_model=StageC) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index b7c3be309..795af828a 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -182,9 +182,9 @@ def detect_unet_config(state_dict, key_prefix): return unet_config -def model_config_from_unet_config(unet_config): +def model_config_from_unet_config(unet_config, state_dict=None): for model_config in comfy.supported_models.models: - if model_config.matches(unet_config): + if model_config.matches(unet_config, state_dict): return model_config(unet_config) logging.error("no match {}".format(unet_config)) @@ -192,7 +192,7 @@ def model_config_from_unet_config(unet_config): def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=False): unet_config = detect_unet_config(state_dict, unet_key_prefix) - model_config = model_config_from_unet_config(unet_config) + model_config = model_config_from_unet_config(unet_config, state_dict) if model_config is None and use_base_if_no_match: return comfy.supported_models_base.BASE(unet_config) else: @@ -321,6 +321,12 @@ def unet_config_from_diffusers_unet(state_dict, dtype=None): 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 2, 2, 2, 10, 10, 10], 'use_temporal_attention': False, 'use_temporal_resblock': False} + SDXL_diffusers_ip2p = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, + 'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 8, 'model_channels': 320, + 'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 10, 10], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 10, + 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 2, 2, 2, 10, 10, 10], + 'use_temporal_attention': False, 'use_temporal_resblock': False} + SSD_1B = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 4, 4], 'transformer_depth_output': [0, 0, 0, 1, 1, 2, 10, 4, 4], @@ -345,7 +351,13 @@ def unet_config_from_diffusers_unet(state_dict, dtype=None): 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 6, 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'use_temporal_attention': False, 'use_temporal_resblock': False} - supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B, Segmind_Vega, KOALA_700M, KOALA_1B] + SD09_XS = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, + 'adm_in_channels': None, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [1, 1, 1], + 'transformer_depth': [1, 1, 1], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': -2, 'use_linear_in_transformer': True, + 'context_dim': 1024, 'num_head_channels': 64, 'transformer_depth_output': [1, 1, 1, 1, 1, 1], + 'use_temporal_attention': False, 'use_temporal_resblock': False, 'disable_self_attentions': [True, False, False]} + + supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B, Segmind_Vega, KOALA_700M, KOALA_1B, SD09_XS, SDXL_diffusers_ip2p] for unet_config in supported_models: matches = True diff --git a/comfy/model_management.py b/comfy/model_management.py index dd262e260..26216432a 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -30,7 +30,7 @@ lowvram_available = True xpu_available = False if args.deterministic: - logging.warning("Using deterministic algorithms for pytorch") + logging.info("Using deterministic algorithms for pytorch") torch.use_deterministic_algorithms(True, warn_only=True) directml_enabled = False @@ -42,7 +42,7 @@ if args.directml is not None: directml_device = torch_directml.device() else: directml_device = torch_directml.device(device_index) - logging.warning("Using directml with device: {}".format(torch_directml.device_name(device_index))) + logging.info("Using directml with device: {}".format(torch_directml.device_name(device_index))) # torch_directml.disable_tiled_resources(True) lowvram_available = False #TODO: need to find a way to get free memory in directml before this can be enabled by default. @@ -118,7 +118,7 @@ def get_total_memory(dev=None, torch_total_too=False): total_vram = get_total_memory(get_torch_device()) / (1024 * 1024) total_ram = psutil.virtual_memory().total / (1024 * 1024) -logging.warning("Total VRAM {:0.0f} MB, total RAM {:0.0f} MB".format(total_vram, total_ram)) +logging.info("Total VRAM {:0.0f} MB, total RAM {:0.0f} MB".format(total_vram, total_ram)) if not args.normalvram and not args.cpu: if lowvram_available and total_vram <= 4096: logging.warning("Trying to enable lowvram mode because your GPU seems to have 4GB or less. If you don't want this use: --normalvram") @@ -144,7 +144,7 @@ else: pass try: XFORMERS_VERSION = xformers.version.__version__ - logging.warning("xformers version: {}".format(XFORMERS_VERSION)) + logging.info("xformers version: {}".format(XFORMERS_VERSION)) if XFORMERS_VERSION.startswith("0.0.18"): logging.warning("\nWARNING: This version of xformers has a major bug where you will get black images when generating high resolution images.") logging.warning("Please downgrade or upgrade xformers to a different version.\n") @@ -212,11 +212,11 @@ elif args.highvram or args.gpu_only: FORCE_FP32 = False FORCE_FP16 = False if args.force_fp32: - logging.warning("Forcing FP32, if this improves things please report it.") + logging.info("Forcing FP32, if this improves things please report it.") FORCE_FP32 = True if args.force_fp16: - logging.warning("Forcing FP16.") + logging.info("Forcing FP16.") FORCE_FP16 = True if lowvram_available: @@ -230,12 +230,12 @@ if cpu_state != CPUState.GPU: if cpu_state == CPUState.MPS: vram_state = VRAMState.SHARED -logging.warning(f"Set vram state to: {vram_state.name}") +logging.info(f"Set vram state to: {vram_state.name}") DISABLE_SMART_MEMORY = args.disable_smart_memory if DISABLE_SMART_MEMORY: - logging.warning("Disabling smart memory management") + logging.info("Disabling smart memory management") def get_torch_device_name(device): if hasattr(device, 'type'): @@ -253,11 +253,11 @@ def get_torch_device_name(device): return "CUDA {}: {}".format(device, torch.cuda.get_device_name(device)) try: - logging.warning("Device: {}".format(get_torch_device_name(get_torch_device()))) + logging.info("Device: {}".format(get_torch_device_name(get_torch_device()))) except: logging.warning("Could not pick default device.") -logging.warning("VAE dtype: {}".format(VAE_DTYPE)) +logging.info("VAE dtype: {}".format(VAE_DTYPE)) current_loaded_models = [] @@ -272,8 +272,9 @@ def module_size(module): class LoadedModel: def __init__(self, model): self.model = model - self.model_accelerated = False self.device = model.load_device + self.weights_loaded = False + self.real_model = None def model_memory(self): return self.model.model_size() @@ -285,54 +286,34 @@ class LoadedModel: return self.model_memory() def model_load(self, lowvram_model_memory=0): - patch_model_to = None - if lowvram_model_memory == 0: - patch_model_to = self.device + patch_model_to = self.device self.model.model_patches_to(self.device) self.model.model_patches_to(self.model.model_dtype()) + load_weights = not self.weights_loaded + try: - self.real_model = self.model.patch_model(device_to=patch_model_to) #TODO: do something with loras and offloading to CPU + if lowvram_model_memory > 0 and load_weights: + self.real_model = self.model.patch_model_lowvram(device_to=patch_model_to, lowvram_model_memory=lowvram_model_memory) + else: + self.real_model = self.model.patch_model(device_to=patch_model_to, patch_weights=load_weights) except Exception as e: self.model.unpatch_model(self.model.offload_device) self.model_unload() raise e - if lowvram_model_memory > 0: - logging.warning("loading in lowvram mode {}".format(lowvram_model_memory/(1024 * 1024))) - mem_counter = 0 - for m in self.real_model.modules(): - if hasattr(m, "comfy_cast_weights"): - m.prev_comfy_cast_weights = m.comfy_cast_weights - m.comfy_cast_weights = True - module_mem = module_size(m) - if mem_counter + module_mem < lowvram_model_memory: - m.to(self.device) - mem_counter += module_mem - elif hasattr(m, "weight"): #only modules with comfy_cast_weights can be set to lowvram mode - m.to(self.device) - mem_counter += module_size(m) - logging.warning("lowvram: loaded module regularly {}".format(m)) - - self.model_accelerated = True - if is_intel_xpu() and not args.disable_ipex_optimize: self.real_model = torch.xpu.optimize(self.real_model.eval(), inplace=True, auto_kernel_selection=True, graph_mode=True) + self.weights_loaded = True return self.real_model - def model_unload(self): - if self.model_accelerated: - for m in self.real_model.modules(): - if hasattr(m, "prev_comfy_cast_weights"): - m.comfy_cast_weights = m.prev_comfy_cast_weights - del m.prev_comfy_cast_weights - - self.model_accelerated = False - - self.model.unpatch_model(self.model.offload_device) + def model_unload(self, unpatch_weights=True): + self.model.unpatch_model(self.model.offload_device, unpatch_weights=unpatch_weights) self.model.model_patches_to(self.model.offload_device) + self.weights_loaded = self.weights_loaded and not unpatch_weights + self.real_model = None def __eq__(self, other): return self.model is other.model @@ -340,31 +321,57 @@ class LoadedModel: def minimum_inference_memory(): return (1024 * 1024 * 1024) -def unload_model_clones(model): +def unload_model_clones(model, unload_weights_only=True, force_unload=True): to_unload = [] for i in range(len(current_loaded_models)): if model.is_clone(current_loaded_models[i].model): to_unload = [i] + to_unload + if len(to_unload) == 0: + return True + + same_weights = 0 for i in to_unload: - logging.warning("unload clone {}".format(i)) - current_loaded_models.pop(i).model_unload() + if model.clone_has_same_weights(current_loaded_models[i].model): + same_weights += 1 + + if same_weights == len(to_unload): + unload_weight = False + else: + unload_weight = True + + if not force_unload: + if unload_weights_only and unload_weight == False: + return None + + for i in to_unload: + logging.debug("unload clone {} {}".format(i, unload_weight)) + current_loaded_models.pop(i).model_unload(unpatch_weights=unload_weight) + + return unload_weight def free_memory(memory_required, device, keep_loaded=[]): - unloaded_model = False + unloaded_model = [] + can_unload = [] + for i in range(len(current_loaded_models) -1, -1, -1): - if not DISABLE_SMART_MEMORY: - if get_free_memory(device) > memory_required: - break shift_model = current_loaded_models[i] if shift_model.device == device: if shift_model not in keep_loaded: - m = current_loaded_models.pop(i) - m.model_unload() - del m - unloaded_model = True + can_unload.append((sys.getrefcount(shift_model.model), shift_model.model_memory(), i)) - if unloaded_model: + for x in sorted(can_unload): + i = x[-1] + if not DISABLE_SMART_MEMORY: + if get_free_memory(device) > memory_required: + break + current_loaded_models[i].model_unload() + unloaded_model.append(i) + + for i in sorted(unloaded_model, reverse=True): + current_loaded_models.pop(i) + + if len(unloaded_model) > 0: soft_empty_cache() else: if vram_state != VRAMState.HIGH_VRAM: @@ -389,7 +396,7 @@ def load_models_gpu(models, memory_required=0): models_already_loaded.append(loaded_model) else: if hasattr(x, "model"): - logging.warning(f"Requested to load {x.model.__class__.__name__}") + logging.info(f"Requested to load {x.model.__class__.__name__}") models_to_load.append(loaded_model) if len(models_to_load) == 0: @@ -399,17 +406,22 @@ def load_models_gpu(models, memory_required=0): free_memory(extra_mem, d, models_already_loaded) return - logging.warning(f"Loading {len(models_to_load)} new model{'s' if len(models_to_load) > 1 else ''}") + logging.info(f"Loading {len(models_to_load)} new model{'s' if len(models_to_load) > 1 else ''}") total_memory_required = {} for loaded_model in models_to_load: - unload_model_clones(loaded_model.model) - total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device) + if unload_model_clones(loaded_model.model, unload_weights_only=True, force_unload=False) == True:#unload clones where the weights are different + total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device) for device in total_memory_required: if device != torch.device("cpu"): free_memory(total_memory_required[device] * 1.3 + extra_mem, device, models_already_loaded) + for loaded_model in models_to_load: + weights_unloaded = unload_model_clones(loaded_model.model, unload_weights_only=False, force_unload=False) #unload the rest of the clones where the weights can stay loaded + if weights_unloaded is not None: + loaded_model.weights_loaded = not weights_unloaded + for loaded_model in models_to_load: model = loaded_model.model torch_dev = model.load_device @@ -438,11 +450,15 @@ def load_models_gpu(models, memory_required=0): def load_model_gpu(model): return load_models_gpu([model]) -def cleanup_models(): +def cleanup_models(keep_clone_weights_loaded=False): to_delete = [] for i in range(len(current_loaded_models)): if sys.getrefcount(current_loaded_models[i].model) <= 2: - to_delete = [i] + to_delete + if not keep_clone_weights_loaded: + to_delete = [i] + to_delete + #TODO: find a less fragile way to do this. + elif sys.getrefcount(current_loaded_models[i].real_model) <= 3: #references from .real_model + the .model + to_delete = [i] + to_delete for i in to_delete: x = current_loaded_models.pop(i) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 5e578dffc..8dda84cfd 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -2,10 +2,23 @@ import torch import copy import inspect import logging +import uuid import comfy.utils import comfy.model_management +def apply_weight_decompose(dora_scale, weight): + weight_norm = ( + weight.transpose(0, 1) + .reshape(weight.shape[1], -1) + .norm(dim=1, keepdim=True) + .reshape(weight.shape[1], *[1] * (weight.dim() - 1)) + .transpose(0, 1) + ) + + return weight * (dora_scale / weight_norm) + + class ModelPatcher: def __init__(self, model, load_device, offload_device, size=0, current_device=None, weight_inplace_update=False): self.size = size @@ -24,6 +37,8 @@ class ModelPatcher: self.current_device = current_device self.weight_inplace_update = weight_inplace_update + self.model_lowvram = False + self.patches_uuid = uuid.uuid4() def model_size(self): if self.size > 0: @@ -38,10 +53,13 @@ class ModelPatcher: n.patches = {} for k in self.patches: n.patches[k] = self.patches[k][:] + n.patches_uuid = self.patches_uuid n.object_patches = self.object_patches.copy() n.model_options = copy.deepcopy(self.model_options) n.model_keys = self.model_keys + n.backup = self.backup + n.object_patches_backup = self.object_patches_backup return n def is_clone(self, other): @@ -49,6 +67,19 @@ class ModelPatcher: return True return False + def clone_has_same_weights(self, clone): + if not self.is_clone(clone): + return False + + if len(self.patches) == 0 and len(clone.patches) == 0: + return True + + if self.patches_uuid == clone.patches_uuid: + if len(self.patches) != len(clone.patches): + logging.warning("WARNING: something went wrong, same patch uuid but different length of patches.") + else: + return True + def memory_required(self, input_shape): return self.model.memory_required(input_shape=input_shape) @@ -153,6 +184,7 @@ class ModelPatcher: current_patches.append((strength_patch, patches[k], strength_model)) self.patches[k] = current_patches + self.patches_uuid = uuid.uuid4() return list(p) def get_key_patches(self, filter_prefix=None): @@ -178,6 +210,27 @@ class ModelPatcher: sd.pop(k) return sd + def patch_weight_to_device(self, key, device_to=None): + if key not in self.patches: + return + + weight = comfy.utils.get_attr(self.model, key) + + inplace_update = self.weight_inplace_update + + if key not in self.backup: + self.backup[key] = weight.to(device=self.offload_device, copy=inplace_update) + + if device_to is not None: + temp_weight = comfy.model_management.cast_to_device(weight, device_to, torch.float32, copy=True) + else: + temp_weight = weight.to(torch.float32, copy=True) + out_weight = self.calculate_weight(self.patches[key], temp_weight, key).to(weight.dtype) + if inplace_update: + comfy.utils.copy_to_param(self.model, key, out_weight) + else: + comfy.utils.set_attr_param(self.model, key, out_weight) + def patch_model(self, device_to=None, patch_weights=True): for k in self.object_patches: old = comfy.utils.set_attr(self.model, k, self.object_patches[k]) @@ -191,23 +244,7 @@ class ModelPatcher: logging.warning("could not patch. key doesn't exist in model: {}".format(key)) continue - weight = model_sd[key] - - inplace_update = self.weight_inplace_update - - if key not in self.backup: - self.backup[key] = weight.to(device=self.offload_device, copy=inplace_update) - - if device_to is not None: - temp_weight = comfy.model_management.cast_to_device(weight, device_to, torch.float32, copy=True) - else: - temp_weight = weight.to(torch.float32, copy=True) - out_weight = self.calculate_weight(self.patches[key], temp_weight, key).to(weight.dtype) - if inplace_update: - comfy.utils.copy_to_param(self.model, key, out_weight) - else: - comfy.utils.set_attr_param(self.model, key, out_weight) - del temp_weight + self.patch_weight_to_device(key, device_to) if device_to is not None: self.model.to(device_to) @@ -215,6 +252,47 @@ class ModelPatcher: return self.model + def patch_model_lowvram(self, device_to=None, lowvram_model_memory=0): + self.patch_model(device_to, patch_weights=False) + + logging.info("loading in lowvram mode {}".format(lowvram_model_memory/(1024 * 1024))) + class LowVramPatch: + def __init__(self, key, model_patcher): + self.key = key + self.model_patcher = model_patcher + def __call__(self, weight): + return self.model_patcher.calculate_weight(self.model_patcher.patches[self.key], weight, self.key) + + mem_counter = 0 + for n, m in self.model.named_modules(): + lowvram_weight = False + if hasattr(m, "comfy_cast_weights"): + module_mem = comfy.model_management.module_size(m) + if mem_counter + module_mem >= lowvram_model_memory: + lowvram_weight = True + + weight_key = "{}.weight".format(n) + bias_key = "{}.bias".format(n) + + if lowvram_weight: + if weight_key in self.patches: + m.weight_function = LowVramPatch(weight_key, self) + if bias_key in self.patches: + m.bias_function = LowVramPatch(weight_key, self) + + m.prev_comfy_cast_weights = m.comfy_cast_weights + m.comfy_cast_weights = True + else: + if hasattr(m, "weight"): + self.patch_weight_to_device(weight_key, device_to) + self.patch_weight_to_device(bias_key, device_to) + m.to(device_to) + mem_counter += comfy.model_management.module_size(m) + logging.debug("lowvram: loaded module regularly {}".format(m)) + + self.model_lowvram = True + return self.model + def calculate_weight(self, patches, weight, key): for p in patches: alpha = p[0] @@ -243,6 +321,7 @@ class ModelPatcher: elif patch_type == "lora": #lora/locon mat1 = comfy.model_management.cast_to_device(v[0], weight.device, torch.float32) mat2 = comfy.model_management.cast_to_device(v[1], weight.device, torch.float32) + dora_scale = v[4] if v[2] is not None: alpha *= v[2] / mat2.shape[0] if v[3] is not None: @@ -252,6 +331,8 @@ class ModelPatcher: mat2 = torch.mm(mat2.transpose(0, 1).flatten(start_dim=1), mat3.transpose(0, 1).flatten(start_dim=1)).reshape(final_shape).transpose(0, 1) try: weight += (alpha * torch.mm(mat1.flatten(start_dim=1), mat2.flatten(start_dim=1))).reshape(weight.shape).type(weight.dtype) + if dora_scale is not None: + weight = apply_weight_decompose(comfy.model_management.cast_to_device(dora_scale, weight.device, torch.float32), weight) except Exception as e: logging.error("ERROR {} {} {}".format(patch_type, key, e)) elif patch_type == "lokr": @@ -262,6 +343,7 @@ class ModelPatcher: w2_a = v[5] w2_b = v[6] t2 = v[7] + dora_scale = v[8] dim = None if w1 is None: @@ -291,6 +373,8 @@ class ModelPatcher: try: weight += alpha * torch.kron(w1, w2).reshape(weight.shape).type(weight.dtype) + if dora_scale is not None: + weight = apply_weight_decompose(comfy.model_management.cast_to_device(dora_scale, weight.device, torch.float32), weight) except Exception as e: logging.error("ERROR {} {} {}".format(patch_type, key, e)) elif patch_type == "loha": @@ -300,6 +384,7 @@ class ModelPatcher: alpha *= v[2] / w1b.shape[0] w2a = v[3] w2b = v[4] + dora_scale = v[7] if v[5] is not None: #cp decomposition t1 = v[5] t2 = v[6] @@ -320,12 +405,16 @@ class ModelPatcher: try: weight += (alpha * m1 * m2).reshape(weight.shape).type(weight.dtype) + if dora_scale is not None: + weight = apply_weight_decompose(comfy.model_management.cast_to_device(dora_scale, weight.device, torch.float32), weight) except Exception as e: logging.error("ERROR {} {} {}".format(patch_type, key, e)) elif patch_type == "glora": if v[4] is not None: alpha *= v[4] / v[0].shape[0] + dora_scale = v[5] + a1 = comfy.model_management.cast_to_device(v[0].flatten(start_dim=1), weight.device, torch.float32) a2 = comfy.model_management.cast_to_device(v[1].flatten(start_dim=1), weight.device, torch.float32) b1 = comfy.model_management.cast_to_device(v[2].flatten(start_dim=1), weight.device, torch.float32) @@ -333,6 +422,8 @@ class ModelPatcher: try: weight += ((torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1), a2), a1)) * alpha).reshape(weight.shape).type(weight.dtype) + if dora_scale is not None: + weight = apply_weight_decompose(comfy.model_management.cast_to_device(dora_scale, weight.device, torch.float32), weight) except Exception as e: logging.error("ERROR {} {} {}".format(patch_type, key, e)) else: @@ -340,21 +431,32 @@ class ModelPatcher: return weight - def unpatch_model(self, device_to=None): - keys = list(self.backup.keys()) + def unpatch_model(self, device_to=None, unpatch_weights=True): + if unpatch_weights: + if self.model_lowvram: + for m in self.model.modules(): + if hasattr(m, "prev_comfy_cast_weights"): + m.comfy_cast_weights = m.prev_comfy_cast_weights + del m.prev_comfy_cast_weights + m.weight_function = None + m.bias_function = None - if self.weight_inplace_update: - for k in keys: - comfy.utils.copy_to_param(self.model, k, self.backup[k]) - else: - for k in keys: - comfy.utils.set_attr_param(self.model, k, self.backup[k]) + self.model_lowvram = False - self.backup = {} + keys = list(self.backup.keys()) - if device_to is not None: - self.model.to(device_to) - self.current_device = device_to + if self.weight_inplace_update: + for k in keys: + comfy.utils.copy_to_param(self.model, k, self.backup[k]) + else: + for k in keys: + comfy.utils.set_attr_param(self.model, k, self.backup[k]) + + self.backup.clear() + + if device_to is not None: + self.model.to(device_to) + self.current_device = device_to keys = list(self.object_patches_backup.keys()) for k in keys: diff --git a/comfy/model_sampling.py b/comfy/model_sampling.py index d325f76d9..37976b326 100644 --- a/comfy/model_sampling.py +++ b/comfy/model_sampling.py @@ -20,6 +20,9 @@ class EPS: noise += latent_image return noise + def inverse_noise_scaling(self, sigma, latent): + return latent + class V_PREDICTION(EPS): def calculate_denoised(self, sigma, model_output, model_input): sigma = sigma.view(sigma.shape[:1] + (1,) * (model_output.ndim - 1)) diff --git a/comfy/ops.py b/comfy/ops.py index 517688e8b..eb6507682 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -24,13 +24,20 @@ def cast_bias_weight(s, input): non_blocking = comfy.model_management.device_supports_non_blocking(input.device) if s.bias is not None: bias = s.bias.to(device=input.device, dtype=input.dtype, non_blocking=non_blocking) + if s.bias_function is not None: + bias = s.bias_function(bias) weight = s.weight.to(device=input.device, dtype=input.dtype, non_blocking=non_blocking) + if s.weight_function is not None: + weight = s.weight_function(weight) return weight, bias +class CastWeightBiasOp: + comfy_cast_weights = False + weight_function = None + bias_function = None class disable_weight_init: - class Linear(torch.nn.Linear): - comfy_cast_weights = False + class Linear(torch.nn.Linear, CastWeightBiasOp): def reset_parameters(self): return None @@ -44,8 +51,7 @@ class disable_weight_init: else: return super().forward(*args, **kwargs) - class Conv2d(torch.nn.Conv2d): - comfy_cast_weights = False + class Conv2d(torch.nn.Conv2d, CastWeightBiasOp): def reset_parameters(self): return None @@ -59,8 +65,7 @@ class disable_weight_init: else: return super().forward(*args, **kwargs) - class Conv3d(torch.nn.Conv3d): - comfy_cast_weights = False + class Conv3d(torch.nn.Conv3d, CastWeightBiasOp): def reset_parameters(self): return None @@ -74,8 +79,7 @@ class disable_weight_init: else: return super().forward(*args, **kwargs) - class GroupNorm(torch.nn.GroupNorm): - comfy_cast_weights = False + class GroupNorm(torch.nn.GroupNorm, CastWeightBiasOp): def reset_parameters(self): return None @@ -90,8 +94,7 @@ class disable_weight_init: return super().forward(*args, **kwargs) - class LayerNorm(torch.nn.LayerNorm): - comfy_cast_weights = False + class LayerNorm(torch.nn.LayerNorm, CastWeightBiasOp): def reset_parameters(self): return None @@ -109,8 +112,7 @@ class disable_weight_init: else: return super().forward(*args, **kwargs) - class ConvTranspose2d(torch.nn.ConvTranspose2d): - comfy_cast_weights = False + class ConvTranspose2d(torch.nn.ConvTranspose2d, CastWeightBiasOp): def reset_parameters(self): return None diff --git a/comfy/samplers.py b/comfy/samplers.py index 6863be4eb..204a98f95 100644 --- a/comfy/samplers.py +++ b/comfy/samplers.py @@ -4,6 +4,7 @@ import torch import collections from comfy import model_management import math +import logging def get_area_and_mult(conds, x_in, timestep_in): area = (x_in.shape[2], x_in.shape[3], 0, 0) @@ -126,30 +127,23 @@ def cond_cat(c_list): return out -def calc_cond_uncond_batch(model, cond, uncond, x_in, timestep, model_options): - out_cond = torch.zeros_like(x_in) - out_count = torch.ones_like(x_in) * 1e-37 - - out_uncond = torch.zeros_like(x_in) - out_uncond_count = torch.ones_like(x_in) * 1e-37 - - COND = 0 - UNCOND = 1 - +def calc_cond_batch(model, conds, x_in, timestep, model_options): + out_conds = [] + out_counts = [] to_run = [] - for x in cond: - p = get_area_and_mult(x, x_in, timestep) - if p is None: - continue - to_run += [(p, COND)] - if uncond is not None: - for x in uncond: - p = get_area_and_mult(x, x_in, timestep) - if p is None: - continue + for i in range(len(conds)): + out_conds.append(torch.zeros_like(x_in)) + out_counts.append(torch.ones_like(x_in) * 1e-37) - to_run += [(p, UNCOND)] + cond = conds[i] + if cond is not None: + for x in cond: + p = get_area_and_mult(x, x_in, timestep) + if p is None: + continue + + to_run += [(p, i)] while len(to_run) > 0: first = to_run[0] @@ -221,22 +215,20 @@ def calc_cond_uncond_batch(model, cond, uncond, x_in, timestep, model_options): output = model_options['model_function_wrapper'](model.apply_model, {"input": input_x, "timestep": timestep_, "c": c, "cond_or_uncond": cond_or_uncond}).chunk(batch_chunks) else: output = model.apply_model(input_x, timestep_, **c).chunk(batch_chunks) - del input_x for o in range(batch_chunks): - if cond_or_uncond[o] == COND: - out_cond[:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += output[o] * mult[o] - out_count[:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += mult[o] - else: - out_uncond[:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += output[o] * mult[o] - out_uncond_count[:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += mult[o] - del mult + cond_index = cond_or_uncond[o] + out_conds[cond_index][:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += output[o] * mult[o] + out_counts[cond_index][:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += mult[o] - out_cond /= out_count - del out_count - out_uncond /= out_uncond_count - del out_uncond_count - return out_cond, out_uncond + for i in range(len(out_conds)): + out_conds[i] /= out_counts[i] + + return out_conds + +def calc_cond_uncond_batch(model, cond, uncond, x_in, timestep, model_options): #TODO: remove + logging.warning("WARNING: The comfy.samplers.calc_cond_uncond_batch function is deprecated please use the calc_cond_batch one instead.") + return tuple(calc_cond_batch(model, [cond, uncond], x_in, timestep, model_options)) #The main sampling function shared by all the samplers #Returns denoised @@ -246,7 +238,13 @@ def sampling_function(model, x, timestep, uncond, cond, cond_scale, model_option else: uncond_ = uncond - cond_pred, uncond_pred = calc_cond_uncond_batch(model, cond, uncond_, x, timestep, model_options) + + conds = [cond, uncond_] + + out = calc_cond_batch(model, conds, x, timestep, model_options) + cond_pred = out[0] + uncond_pred = out[1] + if "sampler_cfg_function" in model_options: args = {"cond": x - cond_pred, "uncond": x - uncond_pred, "cond_scale": cond_scale, "timestep": timestep, "input": x, "sigma": timestep, "cond_denoised": cond_pred, "uncond_denoised": uncond_pred, "model": model, "model_options": model_options} @@ -545,6 +543,7 @@ class KSAMPLER(Sampler): k_callback = lambda x: callback(x["i"], x["denoised"], x["x"], total_steps) samples = self.sampler_function(model_k, noise, sigmas, extra_args=extra_args, callback=k_callback, disable=disable_pbar, **self.extra_options) + samples = model_wrap.inner_model.model_sampling.inverse_noise_scaling(sigmas[-1], samples) return samples @@ -558,11 +557,11 @@ def ksampler(sampler_name, extra_options={}, inpaint_options={}): return k_diffusion_sampling.sample_dpm_fast(model, noise, sigma_min, sigmas[0], total_steps, extra_args=extra_args, callback=callback, disable=disable) sampler_function = dpm_fast_function elif sampler_name == "dpm_adaptive": - def dpm_adaptive_function(model, noise, sigmas, extra_args, callback, disable): + def dpm_adaptive_function(model, noise, sigmas, extra_args, callback, disable, **extra_options): sigma_min = sigmas[-1] if sigma_min == 0: sigma_min = sigmas[-2] - return k_diffusion_sampling.sample_dpm_adaptive(model, noise, sigma_min, sigmas[0], extra_args=extra_args, callback=callback, disable=disable) + return k_diffusion_sampling.sample_dpm_adaptive(model, noise, sigma_min, sigmas[0], extra_args=extra_args, callback=callback, disable=disable, **extra_options) sampler_function = dpm_adaptive_function else: sampler_function = getattr(k_diffusion_sampling, "sample_{}".format(sampler_name)) @@ -625,7 +624,7 @@ def calculate_sigmas_scheduler(model, scheduler_name, steps): elif scheduler_name == "sgm_uniform": sigmas = normal_scheduler(model, steps, sgm=True) else: - print("error invalid scheduler", scheduler_name) + logging.error("error invalid scheduler {}".format(scheduler_name)) return sigmas def sampler_object(name): diff --git a/comfy/sd.py b/comfy/sd.py index 3e4b9e47f..85821120e 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -229,7 +229,7 @@ class VAE: logging.warning("Missing VAE keys {}".format(m)) if len(u) > 0: - logging.info("Leftover VAE keys {}".format(u)) + logging.debug("Leftover VAE keys {}".format(u)) if device is None: device = model_management.vae_device() @@ -397,7 +397,7 @@ def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DI logging.warning("clip missing: {}".format(m)) if len(u) > 0: - logging.info("clip unexpected: {}".format(u)) + logging.debug("clip unexpected: {}".format(u)) return clip def load_gligen(ckpt_path): @@ -538,18 +538,18 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o logging.warning("clip missing: {}".format(m)) if len(u) > 0: - logging.info("clip unexpected {}:".format(u)) + logging.debug("clip unexpected {}:".format(u)) else: logging.warning("no CLIP/text encoder weights in checkpoint, the text encoder model will not be loaded.") left_over = sd.keys() if len(left_over) > 0: - logging.info("left over keys: {}".format(left_over)) + logging.debug("left over keys: {}".format(left_over)) if output_model: model_patcher = comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device(), current_device=inital_load_device) if inital_load_device != torch.device("cpu"): - logging.warning("loaded straight to GPU") + logging.info("loaded straight to GPU") model_management.load_model_gpu(model_patcher) return (model_patcher, clip, vae, clipvision) @@ -589,7 +589,7 @@ def load_unet_state_dict(sd): #load unet in diffusers format model.load_model_weights(new_sd, "") left_over = sd.keys() if len(left_over) > 0: - logging.warning("left over keys in unet: {}".format(left_over)) + logging.info("left over keys in unet: {}".format(left_over)) return comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=offload_device) def load_unet(unet_path): diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 375821032..9bfe3ea1d 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -45,6 +45,11 @@ class SD15(supported_models_base.BASE): return state_dict def process_clip_state_dict_for_saving(self, state_dict): + pop_keys = ["clip_l.transformer.text_projection.weight", "clip_l.logit_scale"] + for p in pop_keys: + if p in state_dict: + state_dict.pop(p) + replace_prefix = {"clip_l.": "cond_stage_model."} return utils.state_dict_prefix_replace(state_dict, replace_prefix) @@ -65,8 +70,8 @@ class SD20(supported_models_base.BASE): def model_type(self, state_dict, prefix=""): if self.unet_config["in_channels"] == 4: #SD2.0 inpainting models are not v prediction k = "{}output_blocks.11.1.transformer_blocks.0.norm1.bias".format(prefix) - out = state_dict[k] - if torch.std(out, unbiased=False) > 0.09: # not sure how well this will actually work. I guess we will find out. + out = state_dict.get(k, None) + if out is not None and torch.std(out, unbiased=False) > 0.09: # not sure how well this will actually work. I guess we will find out. return model_base.ModelType.V_PREDICTION return model_base.ModelType.EPS @@ -279,6 +284,41 @@ class SVD_img2vid(supported_models_base.BASE): def clip_target(self): return None +class SV3D_u(SVD_img2vid): + unet_config = { + "model_channels": 320, + "in_channels": 8, + "use_linear_in_transformer": True, + "transformer_depth": [1, 1, 1, 1, 1, 1, 0, 0], + "context_dim": 1024, + "adm_in_channels": 256, + "use_temporal_attention": True, + "use_temporal_resblock": True + } + + vae_key_prefix = ["conditioner.embedders.1.encoder."] + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.SV3D_u(self, device=device) + return out + +class SV3D_p(SV3D_u): + unet_config = { + "model_channels": 320, + "in_channels": 8, + "use_linear_in_transformer": True, + "transformer_depth": [1, 1, 1, 1, 1, 1, 0, 0], + "context_dim": 1024, + "adm_in_channels": 1280, + "use_temporal_attention": True, + "use_temporal_resblock": True + } + + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.SV3D_p(self, device=device) + return out + class Stable_Zero123(supported_models_base.BASE): unet_config = { "context_dim": 768, @@ -294,6 +334,11 @@ class Stable_Zero123(supported_models_base.BASE): "num_head_channels": -1, } + required_keys = { + "cc_projection.weight": None, + "cc_projection.bias": None, + } + clip_vision_prefix = "cond_stage_model.model.visual." latent_format = latent_formats.SD15 @@ -399,6 +444,33 @@ class Stable_Cascade_B(Stable_Cascade_C): out = model_base.StableCascade_B(self, device=device) return out +class SD15_instructpix2pix(SD15): + unet_config = { + "context_dim": 768, + "model_channels": 320, + "use_linear_in_transformer": False, + "adm_in_channels": None, + "use_temporal_attention": False, + "in_channels": 8, + } + + def get_model(self, state_dict, prefix="", device=None): + return model_base.SD15_instructpix2pix(self, device=device) + +class SDXL_instructpix2pix(SDXL): + unet_config = { + "model_channels": 320, + "use_linear_in_transformer": True, + "transformer_depth": [0, 0, 2, 2, 10, 10], + "context_dim": 2048, + "adm_in_channels": 2816, + "use_temporal_attention": False, + "in_channels": 8, + } + + def get_model(self, state_dict, prefix="", device=None): + return model_base.SDXL_instructpix2pix(self, device=device) + +models = [Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p] -models = [Stable_Zero123, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B] models += [SVD_img2vid] diff --git a/comfy/supported_models_base.py b/comfy/supported_models_base.py index 4d7e25936..6196daabf 100644 --- a/comfy/supported_models_base.py +++ b/comfy/supported_models_base.py @@ -16,6 +16,8 @@ class BASE: "num_head_channels": 64, } + required_keys = {} + clip_prefix = [] clip_vision_prefix = None noise_aug_config = None @@ -28,10 +30,14 @@ class BASE: manual_cast_dtype = None @classmethod - def matches(s, unet_config): + def matches(s, unet_config, state_dict=None): for k in s.unet_config: if k not in unet_config or s.unet_config[k] != unet_config[k]: return False + if state_dict is not None: + for k in s.required_keys: + if k not in state_dict: + return False return True def model_type(self, state_dict, prefix=""): diff --git a/comfy/utils.py b/comfy/utils.py index 8caecd866..ab47b8f28 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -22,7 +22,7 @@ def load_torch_file(ckpt, safe_load=False, device=None): else: pl_sd = torch.load(ckpt, map_location=device, pickle_module=comfy.checkpoint_pickle) if "global_step" in pl_sd: - logging.info(f"Global Step: {pl_sd['global_step']}") + logging.debug(f"Global Step: {pl_sd['global_step']}") if "state_dict" in pl_sd: sd = pl_sd["state_dict"] else: diff --git a/comfy_extras/nodes_custom_sampler.py b/comfy_extras/nodes_custom_sampler.py index 0ad1246a6..72ff7957f 100644 --- a/comfy_extras/nodes_custom_sampler.py +++ b/comfy_extras/nodes_custom_sampler.py @@ -181,6 +181,28 @@ class KSamplerSelect: sampler = comfy.samplers.sampler_object(sampler_name) return (sampler, ) +class SamplerDPMPP_3M_SDE: + @classmethod + def INPUT_TYPES(s): + return {"required": + {"eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "noise_device": (['gpu', 'cpu'], ), + } + } + RETURN_TYPES = ("SAMPLER",) + CATEGORY = "sampling/custom_sampling/samplers" + + FUNCTION = "get_sampler" + + def get_sampler(self, eta, s_noise, noise_device): + if noise_device == 'cpu': + sampler_name = "dpmpp_3m_sde" + else: + sampler_name = "dpmpp_3m_sde_gpu" + sampler = comfy.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise}) + return (sampler, ) + class SamplerDPMPP_2M_SDE: @classmethod def INPUT_TYPES(s): @@ -245,6 +267,49 @@ class SamplerEulerAncestral: sampler = comfy.samplers.ksampler("euler_ancestral", {"eta": eta, "s_noise": s_noise}) return (sampler, ) +class SamplerLMS: + @classmethod + def INPUT_TYPES(s): + return {"required": + {"order": ("INT", {"default": 4, "min": 1, "max": 100}), + } + } + RETURN_TYPES = ("SAMPLER",) + CATEGORY = "sampling/custom_sampling/samplers" + + FUNCTION = "get_sampler" + + def get_sampler(self, order): + sampler = comfy.samplers.ksampler("lms", {"order": order}) + return (sampler, ) + +class SamplerDPMAdaptative: + @classmethod + def INPUT_TYPES(s): + return {"required": + {"order": ("INT", {"default": 3, "min": 2, "max": 3}), + "rtol": ("FLOAT", {"default": 0.05, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "atol": ("FLOAT", {"default": 0.0078, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "h_init": ("FLOAT", {"default": 0.05, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "pcoeff": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "icoeff": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "dcoeff": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "accept_safety": ("FLOAT", {"default": 0.81, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "eta": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + } + } + RETURN_TYPES = ("SAMPLER",) + CATEGORY = "sampling/custom_sampling/samplers" + + FUNCTION = "get_sampler" + + def get_sampler(self, order, rtol, atol, h_init, pcoeff, icoeff, dcoeff, accept_safety, eta, s_noise): + sampler = comfy.samplers.ksampler("dpm_adaptive", {"order": order, "rtol": rtol, "atol": atol, "h_init": h_init, "pcoeff": pcoeff, + "icoeff": icoeff, "dcoeff": dcoeff, "accept_safety": accept_safety, "eta": eta, + "s_noise":s_noise }) + return (sampler, ) + class SamplerCustom: @classmethod def INPUT_TYPES(s): @@ -306,8 +371,11 @@ NODE_CLASS_MAPPINGS = { "SDTurboScheduler": SDTurboScheduler, "KSamplerSelect": KSamplerSelect, "SamplerEulerAncestral": SamplerEulerAncestral, + "SamplerLMS": SamplerLMS, + "SamplerDPMPP_3M_SDE": SamplerDPMPP_3M_SDE, "SamplerDPMPP_2M_SDE": SamplerDPMPP_2M_SDE, "SamplerDPMPP_SDE": SamplerDPMPP_SDE, + "SamplerDPMAdaptative": SamplerDPMAdaptative, "SplitSigmas": SplitSigmas, "FlipSigmas": FlipSigmas, } diff --git a/comfy_extras/nodes_freelunch.py b/comfy_extras/nodes_freelunch.py index 7764aa0b0..6f1d87bf3 100644 --- a/comfy_extras/nodes_freelunch.py +++ b/comfy_extras/nodes_freelunch.py @@ -1,7 +1,7 @@ #code originally taken from: https://github.com/ChenyangSi/FreeU (under MIT License) import torch - +import logging def Fourier_filter(x, threshold, scale): # FFT @@ -49,7 +49,7 @@ class FreeU: try: hsp = Fourier_filter(hsp, threshold=1, scale=scale[1]) except: - print("Device", hsp.device, "does not support the torch.fft functions used in the FreeU node, switching to CPU.") + logging.warning("Device {} does not support the torch.fft functions used in the FreeU node, switching to CPU.".format(hsp.device)) on_cpu_devices[hsp.device] = True hsp = Fourier_filter(hsp.cpu(), threshold=1, scale=scale[1]).to(hsp.device) else: @@ -95,7 +95,7 @@ class FreeU_V2: try: hsp = Fourier_filter(hsp, threshold=1, scale=scale[1]) except: - print("Device", hsp.device, "does not support the torch.fft functions used in the FreeU node, switching to CPU.") + logging.warning("Device {} does not support the torch.fft functions used in the FreeU node, switching to CPU.".format(hsp.device)) on_cpu_devices[hsp.device] = True hsp = Fourier_filter(hsp.cpu(), threshold=1, scale=scale[1]).to(hsp.device) else: diff --git a/comfy_extras/nodes_hypernetwork.py b/comfy_extras/nodes_hypernetwork.py index f692945a8..cafafa6ab 100644 --- a/comfy_extras/nodes_hypernetwork.py +++ b/comfy_extras/nodes_hypernetwork.py @@ -1,6 +1,7 @@ import comfy.utils import folder_paths import torch +import logging def load_hypernetwork_patch(path, strength): sd = comfy.utils.load_torch_file(path, safe_load=True) @@ -23,7 +24,7 @@ def load_hypernetwork_patch(path, strength): } if activation_func not in valid_activation: - print("Unsupported Hypernetwork format, if you report it I might implement it.", path, " ", activation_func, is_layer_norm, use_dropout, activate_output, last_layer_dropout) + logging.error("Unsupported Hypernetwork format, if you report it I might implement it. {} {} {} {} {} {}".format(path, activation_func, is_layer_norm, use_dropout, activate_output, last_layer_dropout)) return None out = {} diff --git a/comfy_extras/nodes_images.py b/comfy_extras/nodes_images.py index 8f638bf8f..af37666b2 100644 --- a/comfy_extras/nodes_images.py +++ b/comfy_extras/nodes_images.py @@ -37,7 +37,7 @@ class RepeatImageBatch: @classmethod def INPUT_TYPES(s): return {"required": { "image": ("IMAGE",), - "amount": ("INT", {"default": 1, "min": 1, "max": 64}), + "amount": ("INT", {"default": 1, "min": 1, "max": 4096}), }} RETURN_TYPES = ("IMAGE",) FUNCTION = "repeat" @@ -52,8 +52,8 @@ class ImageFromBatch: @classmethod def INPUT_TYPES(s): return {"required": { "image": ("IMAGE",), - "batch_index": ("INT", {"default": 0, "min": 0, "max": 63}), - "length": ("INT", {"default": 1, "min": 1, "max": 64}), + "batch_index": ("INT", {"default": 0, "min": 0, "max": 4095}), + "length": ("INT", {"default": 1, "min": 1, "max": 4096}), }} RETURN_TYPES = ("IMAGE",) FUNCTION = "frombatch" diff --git a/comfy_extras/nodes_perpneg.py b/comfy_extras/nodes_perpneg.py index 64bbc1dcd..9e8a218f9 100644 --- a/comfy_extras/nodes_perpneg.py +++ b/comfy_extras/nodes_perpneg.py @@ -10,7 +10,7 @@ class PerpNeg: def INPUT_TYPES(s): return {"required": {"model": ("MODEL", ), "empty_conditioning": ("CONDITIONING", ), - "neg_scale": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0}), + "neg_scale": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.01}), }} RETURN_TYPES = ("MODEL",) FUNCTION = "patch" @@ -31,7 +31,7 @@ class PerpNeg: model_options = args["model_options"] nocond_processed = comfy.samplers.encode_model_conds(model.extra_conds, nocond, x, x.device, "negative") - (noise_pred_nocond, _) = comfy.samplers.calc_cond_uncond_batch(model, nocond_processed, None, x, sigma, model_options) + (noise_pred_nocond,) = comfy.samplers.calc_cond_batch(model, [nocond_processed], x, sigma, model_options) pos = noise_pred_pos - noise_pred_nocond neg = noise_pred_neg - noise_pred_nocond diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index cb5c7d228..0110b472f 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -204,13 +204,13 @@ class Sharpen: "default": 1.0, "min": 0.1, "max": 10.0, - "step": 0.1 + "step": 0.01 }), "alpha": ("FLOAT", { "default": 1.0, "min": 0.0, "max": 5.0, - "step": 0.1 + "step": 0.01 }), }, } diff --git a/comfy_extras/nodes_sag.py b/comfy_extras/nodes_sag.py index bbd380807..69084e91d 100644 --- a/comfy_extras/nodes_sag.py +++ b/comfy_extras/nodes_sag.py @@ -150,7 +150,7 @@ class SelfAttentionGuidance: degraded = create_blur_map(uncond_pred, uncond_attn, sag_sigma, sag_threshold) degraded_noised = degraded + x - uncond_pred # call into the UNet - (sag, _) = comfy.samplers.calc_cond_uncond_batch(model, uncond, None, degraded_noised, sigma, model_options) + (sag,) = comfy.samplers.calc_cond_batch(model, [uncond], degraded_noised, sigma, model_options) return cfg_result + (degraded - sag) * sag_scale m.set_model_sampler_post_cfg_function(post_cfg_function, disable_cfg1_optimization=True) diff --git a/comfy_extras/nodes_stable3d.py b/comfy_extras/nodes_stable3d.py index 4375d8f96..be2e34c28 100644 --- a/comfy_extras/nodes_stable3d.py +++ b/comfy_extras/nodes_stable3d.py @@ -29,8 +29,8 @@ class StableZero123_Conditioning: "width": ("INT", {"default": 256, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}), "height": ("INT", {"default": 256, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}), "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}), - "elevation": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0}), - "azimuth": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0}), + "elevation": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0, "step": 0.1, "round": False}), + "azimuth": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0, "step": 0.1, "round": False}), }} RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT") RETURN_NAMES = ("positive", "negative", "latent") @@ -62,10 +62,10 @@ class StableZero123_Conditioning_Batched: "width": ("INT", {"default": 256, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}), "height": ("INT", {"default": 256, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}), "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}), - "elevation": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0}), - "azimuth": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0}), - "elevation_batch_increment": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0}), - "azimuth_batch_increment": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0}), + "elevation": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0, "step": 0.1, "round": False}), + "azimuth": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0, "step": 0.1, "round": False}), + "elevation_batch_increment": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0, "step": 0.1, "round": False}), + "azimuth_batch_increment": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0, "step": 0.1, "round": False}), }} RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT") RETURN_NAMES = ("positive", "negative", "latent") @@ -95,8 +95,49 @@ class StableZero123_Conditioning_Batched: latent = torch.zeros([batch_size, 4, height // 8, width // 8]) return (positive, negative, {"samples":latent, "batch_index": [0] * batch_size}) +class SV3D_Conditioning: + @classmethod + def INPUT_TYPES(s): + return {"required": { "clip_vision": ("CLIP_VISION",), + "init_image": ("IMAGE",), + "vae": ("VAE",), + "width": ("INT", {"default": 576, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}), + "height": ("INT", {"default": 576, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}), + "video_frames": ("INT", {"default": 21, "min": 1, "max": 4096}), + "elevation": ("FLOAT", {"default": 0.0, "min": -90.0, "max": 90.0, "step": 0.1, "round": False}), + }} + RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT") + RETURN_NAMES = ("positive", "negative", "latent") + + FUNCTION = "encode" + + CATEGORY = "conditioning/3d_models" + + def encode(self, clip_vision, init_image, vae, width, height, video_frames, elevation): + output = clip_vision.encode_image(init_image) + pooled = output.image_embeds.unsqueeze(0) + pixels = comfy.utils.common_upscale(init_image.movedim(-1,1), width, height, "bilinear", "center").movedim(1,-1) + encode_pixels = pixels[:,:,:,:3] + t = vae.encode(encode_pixels) + + azimuth = 0 + azimuth_increment = 360 / (max(video_frames, 2) - 1) + + elevations = [] + azimuths = [] + for i in range(video_frames): + elevations.append(elevation) + azimuths.append(azimuth) + azimuth += azimuth_increment + + positive = [[pooled, {"concat_latent_image": t, "elevation": elevations, "azimuth": azimuths}]] + negative = [[torch.zeros_like(pooled), {"concat_latent_image": torch.zeros_like(t), "elevation": elevations, "azimuth": azimuths}]] + latent = torch.zeros([video_frames, 4, height // 8, width // 8]) + return (positive, negative, {"samples":latent}) + NODE_CLASS_MAPPINGS = { "StableZero123_Conditioning": StableZero123_Conditioning, "StableZero123_Conditioning_Batched": StableZero123_Conditioning_Batched, + "SV3D_Conditioning": SV3D_Conditioning, } diff --git a/comfy_extras/nodes_stable_cascade.py b/comfy_extras/nodes_stable_cascade.py index 7e2d37d22..fcbbeb27f 100644 --- a/comfy_extras/nodes_stable_cascade.py +++ b/comfy_extras/nodes_stable_cascade.py @@ -74,7 +74,7 @@ class StableCascade_StageC_VAEEncode: s = comfy.utils.common_upscale(image.movedim(-1,1), out_width, out_height, "bicubic", "center").movedim(1,-1) c_latent = vae.encode(s[:,:,:,:3]) - b_latent = torch.zeros([c_latent.shape[0], 4, height // 4, width // 4]) + b_latent = torch.zeros([c_latent.shape[0], 4, (height // 8) * 2, (width // 8) * 2]) return ({ "samples": c_latent, }, { diff --git a/comfy_extras/nodes_video_model.py b/comfy_extras/nodes_video_model.py index a52625652..1a0189ed4 100644 --- a/comfy_extras/nodes_video_model.py +++ b/comfy_extras/nodes_video_model.py @@ -79,6 +79,33 @@ class VideoLinearCFGGuidance: m.set_model_sampler_cfg_function(linear_cfg) return (m, ) +class VideoTriangleCFGGuidance: + @classmethod + def INPUT_TYPES(s): + return {"required": { "model": ("MODEL",), + "min_cfg": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.5, "round": 0.01}), + }} + RETURN_TYPES = ("MODEL",) + FUNCTION = "patch" + + CATEGORY = "sampling/video_models" + + def patch(self, model, min_cfg): + def linear_cfg(args): + cond = args["cond"] + uncond = args["uncond"] + cond_scale = args["cond_scale"] + period = 1.0 + values = torch.linspace(0, 1, cond.shape[0], device=cond.device) + values = 2 * (values / period - torch.floor(values / period + 0.5)).abs() + scale = (values * (cond_scale - min_cfg) + min_cfg).reshape((cond.shape[0], 1, 1, 1)) + + return uncond + scale * (cond - uncond) + + m = model.clone() + m.set_model_sampler_cfg_function(linear_cfg) + return (m, ) + class ImageOnlyCheckpointSave(comfy_extras.nodes_model_merging.CheckpointSave): CATEGORY = "_for_testing" @@ -98,6 +125,7 @@ NODE_CLASS_MAPPINGS = { "ImageOnlyCheckpointLoader": ImageOnlyCheckpointLoader, "SVD_img2vid_Conditioning": SVD_img2vid_Conditioning, "VideoLinearCFGGuidance": VideoLinearCFGGuidance, + "VideoTriangleCFGGuidance": VideoTriangleCFGGuidance, "ImageOnlyCheckpointSave": ImageOnlyCheckpointSave, } diff --git a/cuda_malloc.py b/cuda_malloc.py index 70e7ecf9a..eb2857c5f 100644 --- a/cuda_malloc.py +++ b/cuda_malloc.py @@ -47,7 +47,7 @@ blacklist = {"GeForce GTX TITAN X", "GeForce GTX 980", "GeForce GTX 970", "GeFor "Quadro K1200", "Quadro K2200", "Quadro M500", "Quadro M520", "Quadro M600", "Quadro M620", "Quadro M1000", "Quadro M1200", "Quadro M2000", "Quadro M2200", "Quadro M3000", "Quadro M4000", "Quadro M5000", "Quadro M5500", "Quadro M6000", "GeForce MX110", "GeForce MX130", "GeForce 830M", "GeForce 840M", "GeForce GTX 850M", "GeForce GTX 860M", - "GeForce GTX 1650", "GeForce GTX 1630" + "GeForce GTX 1650", "GeForce GTX 1630", "Tesla M4", "Tesla M6", "Tesla M10", "Tesla M40", "Tesla M60" } def cuda_malloc_supported(): diff --git a/custom_nodes/websocket_image_save.py.disabled b/custom_nodes/websocket_image_save.py similarity index 84% rename from custom_nodes/websocket_image_save.py.disabled rename to custom_nodes/websocket_image_save.py index b85a5de8b..5aa573642 100644 --- a/custom_nodes/websocket_image_save.py.disabled +++ b/custom_nodes/websocket_image_save.py @@ -10,10 +10,6 @@ import time #binary images on the websocket with a 8 byte header indicating the type #of binary message (first 4 bytes) and the image format (next 4 bytes). -#The reason this node is disabled by default is because there is a small -#issue when using it with the default ComfyUI web interface: When generating -#batches only the last image will be shown in the UI. - #Note that no metadata will be put in the images saved with this node. class SaveImageWebsocket: @@ -28,7 +24,7 @@ class SaveImageWebsocket: OUTPUT_NODE = True - CATEGORY = "image" + CATEGORY = "api/image" def save_images(self, images): pbar = comfy.utils.ProgressBar(images.shape[0]) diff --git a/execution.py b/execution.py index 5f8ea835c..35bdb77a6 100644 --- a/execution.py +++ b/execution.py @@ -368,7 +368,7 @@ class PromptExecutor: d = self.outputs_ui.pop(x) del d - comfy.model_management.cleanup_models() + comfy.model_management.cleanup_models(keep_clone_weights_loaded=True) self.add_message("execution_cached", { "nodes": list(current_outputs) , "prompt_id": prompt_id}, broadcast=False) diff --git a/folder_paths.py b/folder_paths.py index f1bf40f8c..a84524d9f 100644 --- a/folder_paths.py +++ b/folder_paths.py @@ -1,5 +1,6 @@ import os import time +import logging supported_pt_extensions = set(['.ckpt', '.pt', '.bin', '.pth', '.safetensors']) @@ -44,7 +45,7 @@ if not os.path.exists(input_directory): try: os.makedirs(input_directory) except: - print("Failed to create input directory") + logging.error("Failed to create input directory") def set_output_directory(output_dir): global output_directory @@ -146,21 +147,23 @@ def recursive_search(directory, excluded_dir_names=None): try: dirs[directory] = os.path.getmtime(directory) except FileNotFoundError: - print(f"Warning: Unable to access {directory}. Skipping this path.") - + logging.warning(f"Warning: Unable to access {directory}. Skipping this path.") + + logging.debug("recursive file list on directory {}".format(directory)) for dirpath, subdirs, filenames in os.walk(directory, followlinks=True, topdown=True): subdirs[:] = [d for d in subdirs if d not in excluded_dir_names] for file_name in filenames: relative_path = os.path.relpath(os.path.join(dirpath, file_name), directory) result.append(relative_path) - + for d in subdirs: path = os.path.join(dirpath, d) try: dirs[path] = os.path.getmtime(path) except FileNotFoundError: - print(f"Warning: Unable to access {path}. Skipping this path.") + logging.warning(f"Warning: Unable to access {path}. Skipping this path.") continue + logging.debug("found {} files".format(len(result))) return result, dirs def filter_files_extensions(files, extensions): @@ -248,8 +251,8 @@ def get_save_image_path(filename_prefix, output_dir, image_width=0, image_height err = "**** ERROR: Saving image outside the output folder is not allowed." + \ "\n full_output_folder: " + os.path.abspath(full_output_folder) + \ "\n output_dir: " + output_dir + \ - "\n commonpath: " + os.path.commonpath((output_dir, os.path.abspath(full_output_folder))) - print(err) + "\n commonpath: " + os.path.commonpath((output_dir, os.path.abspath(full_output_folder))) + logging.error(err) raise Exception(err) try: diff --git a/latent_preview.py b/latent_preview.py index 61754751e..4dbcbf455 100644 --- a/latent_preview.py +++ b/latent_preview.py @@ -6,6 +6,7 @@ from comfy.cli_args import args, LatentPreviewMethod from comfy.taesd.taesd import TAESD import folder_paths import comfy.utils +import logging MAX_PREVIEW_RESOLUTION = 512 @@ -70,7 +71,7 @@ def get_previewer(device, latent_format): taesd = TAESD(None, taesd_decoder_path).to(device) previewer = TAESDPreviewerImpl(taesd) else: - print("Warning: TAESD previews enabled, but could not find models/vae_approx/{}".format(latent_format.taesd_decoder_name)) + logging.warning("Warning: TAESD previews enabled, but could not find models/vae_approx/{}".format(latent_format.taesd_decoder_name)) if previewer is None: if latent_format.latent_rgb_factors is not None: diff --git a/main.py b/main.py index 5d07ce2d1..b3a3ebea8 100644 --- a/main.py +++ b/main.py @@ -54,15 +54,15 @@ import threading import gc from comfy.cli_args import args +import logging if os.name == "nt": - import logging logging.getLogger("xformers").addFilter(lambda record: 'A matching Triton is not available' not in record.getMessage()) if __name__ == "__main__": if args.cuda_device is not None: os.environ['CUDA_VISIBLE_DEVICES'] = str(args.cuda_device) - print("Set cuda device to:", args.cuda_device) + logging.info("Set cuda device to: {}".format(args.cuda_device)) if args.deterministic: if 'CUBLAS_WORKSPACE_CONFIG' not in os.environ: @@ -88,7 +88,7 @@ def cuda_malloc_warning(): if b in device_name: cuda_malloc_warning = True if cuda_malloc_warning: - print("\nWARNING: this card most likely does not support cuda-malloc, if you get \"CUDA error\" please run ComfyUI with: --disable-cuda-malloc\n") + logging.warning("\nWARNING: this card most likely does not support cuda-malloc, if you get \"CUDA error\" please run ComfyUI with: --disable-cuda-malloc\n") def prompt_worker(q, server): e = execution.PromptExecutor(server) @@ -121,7 +121,7 @@ def prompt_worker(q, server): current_time = time.perf_counter() execution_time = current_time - execution_start_time - print("Prompt executed in {:.2f} seconds".format(execution_time)) + logging.info("Prompt executed in {:.2f} seconds".format(execution_time)) flags = q.get_flags() free_memory = flags.get("free_memory", False) @@ -139,6 +139,7 @@ def prompt_worker(q, server): if need_gc: current_time = time.perf_counter() if (current_time - last_gc_collect) > gc_collect_interval: + comfy.model_management.cleanup_models() gc.collect() comfy.model_management.soft_empty_cache() last_gc_collect = current_time @@ -182,14 +183,14 @@ def load_extra_path_config(yaml_path): full_path = y if base_path is not None: full_path = os.path.join(base_path, full_path) - print("Adding extra search path", x, full_path) + logging.info("Adding extra search path {} {}".format(x, full_path)) folder_paths.add_model_folder_path(x, full_path) if __name__ == "__main__": if args.temp_directory: temp_dir = os.path.join(os.path.abspath(args.temp_directory), "temp") - print(f"Setting temp directory to: {temp_dir}") + logging.info(f"Setting temp directory to: {temp_dir}") folder_paths.set_temp_directory(temp_dir) cleanup_temp() @@ -224,7 +225,7 @@ if __name__ == "__main__": if args.output_directory: output_dir = os.path.abspath(args.output_directory) - print(f"Setting output directory to: {output_dir}") + logging.info(f"Setting output directory to: {output_dir}") folder_paths.set_output_directory(output_dir) #These are the default folders that checkpoints, clip and vae models will be saved to when using CheckpointSave, etc.. nodes @@ -234,7 +235,7 @@ if __name__ == "__main__": if args.input_directory: input_dir = os.path.abspath(args.input_directory) - print(f"Setting input directory to: {input_dir}") + logging.info(f"Setting input directory to: {input_dir}") folder_paths.set_input_directory(input_dir) if args.quick_test_for_ci: @@ -252,6 +253,6 @@ if __name__ == "__main__": try: loop.run_until_complete(run(server, address=args.listen, port=args.port, verbose=not args.dont_print_server, call_on_start=call_on_start)) except KeyboardInterrupt: - print("\nStopped server") + logging.info("\nStopped server") cleanup_temp() diff --git a/nodes.py b/nodes.py index c334cdb95..5cf020df2 100644 --- a/nodes.py +++ b/nodes.py @@ -8,6 +8,7 @@ import traceback import math import time import random +import logging from PIL import Image, ImageOps, ImageSequence from PIL.PngImagePlugin import PngInfo @@ -40,7 +41,7 @@ def before_node_execution(): def interrupt_processing(value=True): comfy.model_management.interrupt_current_processing(value) -MAX_RESOLUTION=8192 +MAX_RESOLUTION=16384 class CLIPTextEncode: @classmethod @@ -83,7 +84,7 @@ class ConditioningAverage : out = [] if len(conditioning_from) > 1: - print("Warning: ConditioningAverage conditioning_from contains more than 1 cond, only the first one will actually be applied to conditioning_to.") + logging.warning("Warning: ConditioningAverage conditioning_from contains more than 1 cond, only the first one will actually be applied to conditioning_to.") cond_from = conditioning_from[0][0] pooled_output_from = conditioning_from[0][1].get("pooled_output", None) @@ -122,7 +123,7 @@ class ConditioningConcat: out = [] if len(conditioning_from) > 1: - print("Warning: ConditioningConcat conditioning_from contains more than 1 cond, only the first one will actually be applied to conditioning_to.") + logging.warning("Warning: ConditioningConcat conditioning_from contains more than 1 cond, only the first one will actually be applied to conditioning_to.") cond_from = conditioning_from[0][0] @@ -1875,6 +1876,7 @@ def load_custom_node(module_path, ignore=set()): sp = os.path.splitext(module_path) module_name = sp[0] try: + logging.debug("Trying to load custom node {}".format(module_path)) if os.path.isfile(module_path): module_spec = importlib.util.spec_from_file_location(module_name, module_path) module_dir = os.path.split(module_path)[0] @@ -1899,11 +1901,11 @@ def load_custom_node(module_path, ignore=set()): NODE_DISPLAY_NAME_MAPPINGS.update(module.NODE_DISPLAY_NAME_MAPPINGS) return True else: - print(f"Skip {module_path} module for custom nodes due to the lack of NODE_CLASS_MAPPINGS.") + logging.warning(f"Skip {module_path} module for custom nodes due to the lack of NODE_CLASS_MAPPINGS.") return False except Exception as e: - print(traceback.format_exc()) - print(f"Cannot import {module_path} module for custom nodes:", e) + logging.warning(traceback.format_exc()) + logging.warning(f"Cannot import {module_path} module for custom nodes: {e}") return False def load_custom_nodes(): @@ -1924,14 +1926,14 @@ def load_custom_nodes(): node_import_times.append((time.perf_counter() - time_before, module_path, success)) if len(node_import_times) > 0: - print("\nImport times for custom nodes:") + logging.info("\nImport times for custom nodes:") for n in sorted(node_import_times): if n[2]: import_message = "" else: import_message = " (IMPORT FAILED)" - print("{:6.1f} seconds{}:".format(n[0], import_message), n[1]) - print() + logging.info("{:6.1f} seconds{}: {}".format(n[0], import_message, n[1])) + logging.info("") def init_custom_nodes(): extras_dir = os.path.join(os.path.dirname(os.path.realpath(__file__)), "comfy_extras") @@ -1973,12 +1975,12 @@ def init_custom_nodes(): load_custom_nodes() if len(import_failed) > 0: - print("WARNING: some comfy_extras/ nodes did not import correctly. This may be because they are missing some dependencies.\n") + logging.warning("WARNING: some comfy_extras/ nodes did not import correctly. This may be because they are missing some dependencies.\n") for node in import_failed: - print("IMPORT FAILED: {}".format(node)) - print("\nThis issue might be caused by missing newly dependencies added the last time you updated ComfyUI.") + logging.warning("IMPORT FAILED: {}".format(node)) + logging.warning("\nThis issue might be caused by new missing dependencies added the last time you updated ComfyUI.") if args.windows_standalone_build: - print("Please run the update script: update/update_comfyui.bat") + logging.warning("Please run the update script: update/update_comfyui.bat") else: - print("Please do a: pip install -r requirements.txt") - print() + logging.warning("Please do a: pip install -r requirements.txt") + logging.warning("") diff --git a/script_examples/websockets_api_example_ws_images.py b/script_examples/websockets_api_example_ws_images.py new file mode 100644 index 000000000..737488621 --- /dev/null +++ b/script_examples/websockets_api_example_ws_images.py @@ -0,0 +1,159 @@ +#This is an example that uses the websockets api and the SaveImageWebsocket node to get images directly without +#them being saved to disk + +import websocket #NOTE: websocket-client (https://github.com/websocket-client/websocket-client) +import uuid +import json +import urllib.request +import urllib.parse + +server_address = "127.0.0.1:8188" +client_id = str(uuid.uuid4()) + +def queue_prompt(prompt): + p = {"prompt": prompt, "client_id": client_id} + data = json.dumps(p).encode('utf-8') + req = urllib.request.Request("http://{}/prompt".format(server_address), data=data) + return json.loads(urllib.request.urlopen(req).read()) + +def get_image(filename, subfolder, folder_type): + data = {"filename": filename, "subfolder": subfolder, "type": folder_type} + url_values = urllib.parse.urlencode(data) + with urllib.request.urlopen("http://{}/view?{}".format(server_address, url_values)) as response: + return response.read() + +def get_history(prompt_id): + with urllib.request.urlopen("http://{}/history/{}".format(server_address, prompt_id)) as response: + return json.loads(response.read()) + +def get_images(ws, prompt): + prompt_id = queue_prompt(prompt)['prompt_id'] + output_images = {} + current_node = "" + while True: + out = ws.recv() + if isinstance(out, str): + message = json.loads(out) + if message['type'] == 'executing': + data = message['data'] + if data['prompt_id'] == prompt_id: + if data['node'] is None: + break #Execution is done + else: + current_node = data['node'] + else: + if current_node == 'save_image_websocket_node': + images_output = output_images.get(current_node, []) + images_output.append(out[8:]) + output_images[current_node] = images_output + + return output_images + +prompt_text = """ +{ + "3": { + "class_type": "KSampler", + "inputs": { + "cfg": 8, + "denoise": 1, + "latent_image": [ + "5", + 0 + ], + "model": [ + "4", + 0 + ], + "negative": [ + "7", + 0 + ], + "positive": [ + "6", + 0 + ], + "sampler_name": "euler", + "scheduler": "normal", + "seed": 8566257, + "steps": 20 + } + }, + "4": { + "class_type": "CheckpointLoaderSimple", + "inputs": { + "ckpt_name": "v1-5-pruned-emaonly.ckpt" + } + }, + "5": { + "class_type": "EmptyLatentImage", + "inputs": { + "batch_size": 1, + "height": 512, + "width": 512 + } + }, + "6": { + "class_type": "CLIPTextEncode", + "inputs": { + "clip": [ + "4", + 1 + ], + "text": "masterpiece best quality girl" + } + }, + "7": { + "class_type": "CLIPTextEncode", + "inputs": { + "clip": [ + "4", + 1 + ], + "text": "bad hands" + } + }, + "8": { + "class_type": "VAEDecode", + "inputs": { + "samples": [ + "3", + 0 + ], + "vae": [ + "4", + 2 + ] + } + }, + "save_image_websocket_node": { + "class_type": "SaveImageWebsocket", + "inputs": { + "images": [ + "8", + 0 + ] + } + } +} +""" + +prompt = json.loads(prompt_text) +#set the text prompt for our positive CLIPTextEncode +prompt["6"]["inputs"]["text"] = "masterpiece best quality man" + +#set the seed for our KSampler node +prompt["3"]["inputs"]["seed"] = 5 + +ws = websocket.WebSocket() +ws.connect("ws://{}/ws?clientId={}".format(server_address, client_id)) +images = get_images(ws, prompt) + +#Commented out code to display the output images: + +# for node_id in images: +# for image_data in images[node_id]: +# from PIL import Image +# import io +# image = Image.open(io.BytesIO(image_data)) +# image.show() + diff --git a/server.py b/server.py index c6132cdf9..5642bd5e2 100644 --- a/server.py +++ b/server.py @@ -15,15 +15,9 @@ from PIL import Image, ImageOps from PIL.PngImagePlugin import PngInfo from io import BytesIO -try: - import aiohttp - from aiohttp import web -except ImportError: - print("Module 'aiohttp' not installed. Please install it via:") - print("pip install aiohttp") - print("or") - print("pip install -r requirements.txt") - sys.exit() +import aiohttp +from aiohttp import web +import logging import mimetypes from comfy.cli_args import args @@ -40,7 +34,7 @@ async def send_socket_catch_exception(function, message): try: await function(message) except (aiohttp.ClientError, aiohttp.ClientPayloadError, ConnectionResetError) as err: - print("send error:", err) + logging.warning("send error: {}".format(err)) @web.middleware async def cache_control(request: web.Request, handler): @@ -118,7 +112,7 @@ class PromptServer(): async for msg in ws: if msg.type == aiohttp.WSMsgType.ERROR: - print('ws connection closed with exception %s' % ws.exception()) + logging.warning('ws connection closed with exception %s' % ws.exception()) finally: self.sockets.pop(sid, None) return ws @@ -419,8 +413,8 @@ class PromptServer(): try: out[x] = node_info(x) except Exception as e: - print(f"[ERROR] An error occurred while retrieving information for the '{x}' node.", file=sys.stderr) - traceback.print_exc() + logging.error(f"[ERROR] An error occurred while retrieving information for the '{x}' node.") + logging.error(traceback.format_exc()) return web.json_response(out) @routes.get("/object_info/{node_class}") @@ -453,7 +447,7 @@ class PromptServer(): @routes.post("/prompt") async def post_prompt(request): - print("got prompt") + logging.info("got prompt") resp_code = 200 out_string = "" json_data = await request.json() @@ -485,7 +479,7 @@ class PromptServer(): response = {"prompt_id": prompt_id, "number": number, "node_errors": valid[3]} return web.json_response(response) else: - print("invalid prompt:", valid[1]) + logging.warning("invalid prompt: {}".format(valid[1])) return web.json_response({"error": valid[1], "node_errors": valid[3]}, status=400) else: return web.json_response({"error": "no prompt", "node_errors": []}, status=400) @@ -633,8 +627,8 @@ class PromptServer(): await site.start() if verbose: - print("Starting server\n") - print("To see the GUI go to: http://{}:{}".format(address, port)) + logging.info("Starting server\n") + logging.info("To see the GUI go to: http://{}:{}".format(address, port)) if call_on_start is not None: call_on_start(address, port) @@ -646,7 +640,7 @@ class PromptServer(): try: json_data = handler(json_data) except Exception as e: - print(f"[ERROR] An error occurred during the on_prompt_handler processing") - traceback.print_exc() + logging.warning(f"[ERROR] An error occurred during the on_prompt_handler processing") + logging.warning(traceback.format_exc()) return json_data diff --git a/tests-ui/tests/groupNode.test.js b/tests-ui/tests/groupNode.test.js index e6ebedd91..53a5828d1 100644 --- a/tests-ui/tests/groupNode.test.js +++ b/tests-ui/tests/groupNode.test.js @@ -947,7 +947,7 @@ describe("group node", () => { expect(p1.widgets.value.widget.options?.step).toBe(80); // width/height step * 10 expect(p2.widgets.value.widget.options?.min).toBe(16); // width/height min - expect(p2.widgets.value.widget.options?.max).toBe(8192); // width/height max + expect(p2.widgets.value.widget.options?.max).toBe(16384); // width/height max expect(p2.widgets.value.widget.options?.step).toBe(80); // width/height step * 10 expect(p1.widgets.value.value).toBe(128);