diff --git a/README.md b/README.md index ba1e844b3..8ca7e4376 100644 --- a/README.md +++ b/README.md @@ -99,7 +99,7 @@ Put your VAE in: models/vae ### AMD GPUs (Linux only) AMD users can install rocm and pytorch with pip if you don't have it already installed, this is the command to install the stable version: -```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.7``` +```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.0``` This is the command to install the nightly with ROCm 6.0 which might have some performance improvements: diff --git a/comfy/samplers.py b/comfy/samplers.py index 415a35cc3..b12b0fd1b 100644 --- a/comfy/samplers.py +++ b/comfy/samplers.py @@ -34,7 +34,7 @@ def get_area_and_mult(conds, x_in, timestep_in): mask = conds['mask'] assert(mask.shape[1] == x_in.shape[2]) assert(mask.shape[2] == x_in.shape[3]) - mask = mask[:,area[2]:area[0] + area[2],area[3]:area[1] + area[3]] * mask_strength + mask = mask[:input_x.shape[0],area[2]:area[0] + area[2],area[3]:area[1] + area[3]] * mask_strength mask = mask.unsqueeze(1).repeat(input_x.shape[0] // mask.shape[0], input_x.shape[1], 1, 1) else: mask = torch.ones_like(input_x) diff --git a/comfy/sd.py b/comfy/sd.py index 3919f4bfa..16dc0b732 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -210,16 +210,26 @@ class VAE: self.first_stage_model = StageC_coder() self.downscale_ratio = 32 self.latent_channels = 16 - else: + elif "decoder.conv_in.weight" in sd: #default SD1.x/SD2.x VAE parameters ddconfig = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0} - if 'encoder.down.2.downsample.conv.weight' not in sd: #Stable diffusion x4 upscaler VAE + if 'encoder.down.2.downsample.conv.weight' not in sd and 'decoder.up.3.upsample.conv.weight' not in sd: #Stable diffusion x4 upscaler VAE ddconfig['ch_mult'] = [1, 2, 4] self.downscale_ratio = 4 self.upscale_ratio = 4 - self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=4) + self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1] + if 'quant_conv.weight' in sd: + self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=4) + else: + self.first_stage_model = AutoencodingEngine(regularizer_config={'target': "comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer"}, + encoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Encoder", 'params': ddconfig}, + decoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Decoder", 'params': ddconfig}) + else: + logging.warning("WARNING: No VAE weights detected, VAE not initalized.") + self.first_stage_model = None + return else: self.first_stage_model = AutoencoderKL(**(config['params'])) self.first_stage_model = self.first_stage_model.eval() diff --git a/comfy_extras/nodes_align_your_steps.py b/comfy_extras/nodes_align_your_steps.py new file mode 100644 index 000000000..3ffe53187 --- /dev/null +++ b/comfy_extras/nodes_align_your_steps.py @@ -0,0 +1,53 @@ +#from: https://research.nvidia.com/labs/toronto-ai/AlignYourSteps/howto.html +import numpy as np +import torch + +def loglinear_interp(t_steps, num_steps): + """ + Performs log-linear interpolation of a given array of decreasing numbers. + """ + xs = np.linspace(0, 1, len(t_steps)) + ys = np.log(t_steps[::-1]) + + new_xs = np.linspace(0, 1, num_steps) + new_ys = np.interp(new_xs, xs, ys) + + interped_ys = np.exp(new_ys)[::-1].copy() + return interped_ys + +NOISE_LEVELS = {"SD1": [14.6146412293, 6.4745760956, 3.8636745985, 2.6946151520, 1.8841921177, 1.3943805092, 0.9642583904, 0.6523686016, 0.3977456272, 0.1515232662, 0.0291671582], + "SDXL":[14.6146412293, 6.3184485287, 3.7681790315, 2.1811480769, 1.3405244945, 0.8620721141, 0.5550693289, 0.3798540708, 0.2332364134, 0.1114188177, 0.0291671582], + "SVD": [700.00, 54.5, 15.886, 7.977, 4.248, 1.789, 0.981, 0.403, 0.173, 0.034, 0.002]} + +class AlignYourStepsScheduler: + @classmethod + def INPUT_TYPES(s): + return {"required": + {"model_type": (["SD1", "SDXL", "SVD"], ), + "steps": ("INT", {"default": 10, "min": 10, "max": 10000}), + "denoise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}), + } + } + RETURN_TYPES = ("SIGMAS",) + CATEGORY = "sampling/custom_sampling/schedulers" + + FUNCTION = "get_sigmas" + + def get_sigmas(self, model_type, steps, denoise): + total_steps = steps + if denoise < 1.0: + if denoise <= 0.0: + return (torch.FloatTensor([]),) + total_steps = round(steps * denoise) + + sigmas = NOISE_LEVELS[model_type][:] + if (steps + 1) != len(sigmas): + sigmas = loglinear_interp(sigmas, steps + 1) + + sigmas = sigmas[-(total_steps + 1):] + sigmas[-1] = 0 + return (torch.FloatTensor(sigmas), ) + +NODE_CLASS_MAPPINGS = { + "AlignYourStepsScheduler": AlignYourStepsScheduler, +} diff --git a/comfy_extras/nodes_custom_sampler.py b/comfy_extras/nodes_custom_sampler.py index 06238f892..7afdbf4bf 100644 --- a/comfy_extras/nodes_custom_sampler.py +++ b/comfy_extras/nodes_custom_sampler.py @@ -39,8 +39,8 @@ class KarrasScheduler: def INPUT_TYPES(s): return {"required": {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}), - "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}), + "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), + "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), "rho": ("FLOAT", {"default": 7.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), } } @@ -58,8 +58,8 @@ class ExponentialScheduler: def INPUT_TYPES(s): return {"required": {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}), - "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}), + "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), + "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), } } RETURN_TYPES = ("SIGMAS",) @@ -76,8 +76,8 @@ class PolyexponentialScheduler: def INPUT_TYPES(s): return {"required": {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}), - "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}), + "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), + "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), "rho": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), } } @@ -117,8 +117,8 @@ class VPScheduler: def INPUT_TYPES(s): return {"required": {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "beta_d": ("FLOAT", {"default": 19.9, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}), #TODO: fix default values - "beta_min": ("FLOAT", {"default": 0.1, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}), + "beta_d": ("FLOAT", {"default": 19.9, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), #TODO: fix default values + "beta_min": ("FLOAT", {"default": 0.1, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), "eps_s": ("FLOAT", {"default": 0.001, "min": 0.0, "max": 1.0, "step":0.0001, "round": False}), } } diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index 0110b472f..68f6ef51e 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -5,6 +5,7 @@ from PIL import Image import math import comfy.utils +import comfy.model_management class Blend: @@ -102,6 +103,7 @@ class Blur: if blur_radius == 0: return (image,) + image = image.to(comfy.model_management.get_torch_device()) batch_size, height, width, channels = image.shape kernel_size = blur_radius * 2 + 1 @@ -112,7 +114,7 @@ class Blur: blurred = F.conv2d(padded_image, kernel, padding=kernel_size // 2, groups=channels)[:,:,blur_radius:-blur_radius, blur_radius:-blur_radius] blurred = blurred.permute(0, 2, 3, 1) - return (blurred,) + return (blurred.to(comfy.model_management.intermediate_device()),) class Quantize: def __init__(self): @@ -225,6 +227,7 @@ class Sharpen: return (image,) batch_size, height, width, channels = image.shape + image = image.to(comfy.model_management.get_torch_device()) kernel_size = sharpen_radius * 2 + 1 kernel = gaussian_kernel(kernel_size, sigma, device=image.device) * -(alpha*10) @@ -239,7 +242,7 @@ class Sharpen: result = torch.clamp(sharpened, 0, 1) - return (result,) + return (result.to(comfy.model_management.intermediate_device()),) class ImageScaleToTotalPixels: upscale_methods = ["nearest-exact", "bilinear", "area", "bicubic", "lanczos"] diff --git a/comfy_extras/nodes_upscale_model.py b/comfy_extras/nodes_upscale_model.py index 2b5e49a55..52c95df23 100644 --- a/comfy_extras/nodes_upscale_model.py +++ b/comfy_extras/nodes_upscale_model.py @@ -37,9 +37,14 @@ class ImageUpscaleWithModel: def upscale(self, upscale_model, image): device = model_management.get_torch_device() + + memory_required = model_management.module_size(upscale_model) + memory_required += (512 * 512 * 3) * image.element_size() * max(upscale_model.scale, 1.0) * 384.0 #The 384.0 is an estimate of how much some of these models take, TODO: make it more accurate + memory_required += image.nelement() * image.element_size() + model_management.free_memory(memory_required, device) + upscale_model.to(device) in_img = image.movedim(-1,-3).to(device) - free_memory = model_management.get_free_memory(device) tile = 512 overlap = 32 diff --git a/execution.py b/execution.py index 35bdb77a6..47d58b9d7 100644 --- a/execution.py +++ b/execution.py @@ -176,7 +176,7 @@ def recursive_execute(server, prompt, outputs, current_item, extra_data, execute for node_id, node_outputs in outputs.items(): output_data_formatted[node_id] = [[format_value(x) for x in l] for l in node_outputs] - logging.error("!!! Exception during processing !!!") + logging.error(f"!!! Exception during processing!!! {ex}") logging.error(traceback.format_exc()) error_details = { diff --git a/folder_paths.py b/folder_paths.py index 3f9a32b4d..489795002 100644 --- a/folder_paths.py +++ b/folder_paths.py @@ -2,7 +2,7 @@ import os import time import logging -supported_pt_extensions = set(['.ckpt', '.pt', '.bin', '.pth', '.safetensors']) +supported_pt_extensions = set(['.ckpt', '.pt', '.bin', '.pth', '.safetensors', '.pkl']) folder_names_and_paths = {} diff --git a/nodes.py b/nodes.py index ea1e32030..a1cfd6365 100644 --- a/nodes.py +++ b/nodes.py @@ -583,8 +583,8 @@ class LoraLoader: return {"required": { "model": ("MODEL",), "clip": ("CLIP", ), "lora_name": (folder_paths.get_filename_list("loras"), ), - "strength_model": ("FLOAT", {"default": 1.0, "min": -20.0, "max": 20.0, "step": 0.01}), - "strength_clip": ("FLOAT", {"default": 1.0, "min": -20.0, "max": 20.0, "step": 0.01}), + "strength_model": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01}), + "strength_clip": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01}), }} RETURN_TYPES = ("MODEL", "CLIP") FUNCTION = "load_lora" @@ -617,7 +617,7 @@ class LoraLoaderModelOnly(LoraLoader): def INPUT_TYPES(s): return {"required": { "model": ("MODEL",), "lora_name": (folder_paths.get_filename_list("loras"), ), - "strength_model": ("FLOAT", {"default": 1.0, "min": -20.0, "max": 20.0, "step": 0.01}), + "strength_model": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01}), }} RETURN_TYPES = ("MODEL",) FUNCTION = "load_lora_model_only" @@ -1943,6 +1943,7 @@ def init_custom_nodes(): "nodes_ip2p.py", "nodes_model_merging_model_specific.py", "nodes_pag.py", + "nodes_align_your_steps.py", ] import_failed = []