Merge branch 'master' of https://github.com/comfyanonymous/ComfyUI

.github/workflows/stable-release.yml

@@ -91,6 +91,8 @@ jobs:
           cd ComfyUI_windows_portable
           python_embeded/python.exe -s ComfyUI/main.py --quick-test-for-ci --cpu
 
+          python_embeded/python.exe -s ./update/update.py ComfyUI/
+
           ls
 
       - name: Upload binaries to release

.github/workflows/windows_release_package.yml

@@ -88,6 +88,8 @@ jobs:
             cd ComfyUI_windows_portable
             python_embeded/python.exe -s ComfyUI/main.py --quick-test-for-ci --cpu
 
+            python_embeded/python.exe -s ./update/update.py ComfyUI/
+
             ls
 
         - name: Upload binaries to release

README.md

@@ -49,7 +49,6 @@ Supports all operating systems and GPU types (NVIDIA, AMD, Intel, Apple Silicon,
 ## [Examples](https://comfyanonymous.github.io/ComfyUI_examples/)
 See what ComfyUI can do with the [example workflows](https://comfyanonymous.github.io/ComfyUI_examples/).
 
-
 ## Features
 - Nodes/graph/flowchart interface to experiment and create complex Stable Diffusion workflows without needing to code anything.
 - Image Models
@@ -99,6 +98,23 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
 
 Workflow examples can be found on the [Examples page](https://comfyanonymous.github.io/ComfyUI_examples/)
 
+## Release Process
+
+ComfyUI follows a weekly release cycle every Friday, with three interconnected repositories:
+
+1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)**
+   - Releases a new stable version (e.g., v0.7.0)
+   - Serves as the foundation for the desktop release
+
+2. **[ComfyUI Desktop](https://github.com/Comfy-Org/desktop)**
+   - Builds a new release using the latest stable core version
+   - Version numbers match the core release (e.g., Desktop v1.7.0 uses Core v1.7.0)
+
+3. **[ComfyUI Frontend](https://github.com/Comfy-Org/ComfyUI_frontend)**
+   - Weekly frontend updates are merged into the core repository
+   - Features are frozen for the upcoming core release
+   - Development continues for the next release cycle
+
 ## Shortcuts
 
 | Keybind                            | Explanation                                                                                                        |
@@ -149,8 +165,6 @@ Simply download, extract with [7-Zip](https://7-zip.org) and run. Make sure you
 
 If you have trouble extracting it, right click the file -> properties -> unblock
 
-If you have a 50 series Blackwell card like a 5090 or 5080 see [this discussion thread](https://github.com/comfyanonymous/ComfyUI/discussions/6643)
-
 #### How do I share models between another UI and ComfyUI?
 
 See the [Config file](extra_model_paths.yaml.example) to set the search paths for models. In the standalone windows build you can find this file in the ComfyUI directory. Rename this file to extra_model_paths.yaml and edit it with your favorite text editor.

app/custom_node_manager.py

@@ -93,16 +93,20 @@ class CustomNodeManager:
 
     def add_routes(self, routes, webapp, loadedModules):
 
+        example_workflow_folder_names = ["example_workflows", "example", "examples", "workflow", "workflows"]
+
         @routes.get("/workflow_templates")
         async def get_workflow_templates(request):
             """Returns a web response that contains the map of custom_nodes names and their associated workflow templates. The ones without templates are omitted."""
-            files = [
+
-                file
+            files = []
-                for folder in folder_paths.get_folder_paths("custom_nodes")
+
-                for file in glob.glob(
+            for folder in folder_paths.get_folder_paths("custom_nodes"):
-                    os.path.join(folder, "*/example_workflows/*.json")
+                for folder_name in example_workflow_folder_names:
-                )
+                    pattern = os.path.join(folder, f"*/{folder_name}/*.json")
-            ]
+                    matched_files = glob.glob(pattern)
+                    files.extend(matched_files)
+
             workflow_templates_dict = (
                 {}
             )  # custom_nodes folder name -> example workflow names
@@ -118,15 +122,22 @@ class CustomNodeManager:
 
         # Serve workflow templates from custom nodes.
         for module_name, module_dir in loadedModules:
-            workflows_dir = os.path.join(module_dir, "example_workflows")
+            for folder_name in example_workflow_folder_names:
-            if os.path.exists(workflows_dir):
+                workflows_dir = os.path.join(module_dir, folder_name)
-                webapp.add_routes(
+
-                    [
+                if os.path.exists(workflows_dir):
-                        web.static(
+                    if folder_name != "example_workflows":
-                            "/api/workflow_templates/" + module_name, workflows_dir
+                        logging.warning(
-                        )
+                            "WARNING: Found example workflow folder '%s' for custom node '%s', consider renaming it to 'example_workflows'",
-                    ]
+                            folder_name, module_name)
-                )
+
+                    webapp.add_routes(
+                        [
+                            web.static(
+                                "/api/workflow_templates/" + module_name, workflows_dir
+                            )
+                        ]
+                    )
 
         @routes.get("/i18n")
         async def get_i18n(request):

comfy/comfy_types/node_typing.py

@@ -48,6 +48,7 @@ class IO(StrEnum):
     FACE_ANALYSIS = "FACE_ANALYSIS"
     BBOX = "BBOX"
     SEGS = "SEGS"
+    VIDEO = "VIDEO"
 
     ANY = "*"
     """Always matches any type, but at a price.
@@ -273,7 +274,7 @@ class ComfyNodeABC(ABC):
 
     Comfy Docs: https://docs.comfy.org/custom-nodes/backend/lists#list-processing
     """
-    OUTPUT_IS_LIST: tuple[bool]
+    OUTPUT_IS_LIST: tuple[bool, ...]
     """A tuple indicating which node outputs are lists, but will be connected to nodes that expect individual items.
 
     Connected nodes that do not implement `INPUT_IS_LIST` will be executed once for every item in the list.
@@ -292,7 +293,7 @@ class ComfyNodeABC(ABC):
     Comfy Docs: https://docs.comfy.org/custom-nodes/backend/lists#list-processing
     """
 
-    RETURN_TYPES: tuple[IO]
+    RETURN_TYPES: tuple[IO, ...]
     """A tuple representing the outputs of this node.
 
     Usage::
@@ -301,12 +302,12 @@ class ComfyNodeABC(ABC):
 
     Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#return-types
     """
-    RETURN_NAMES: tuple[str]
+    RETURN_NAMES: tuple[str, ...]
     """The output slot names for each item in `RETURN_TYPES`, e.g. ``RETURN_NAMES = ("count", "filter_string")``
 
     Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#return-names
     """
-    OUTPUT_TOOLTIPS: tuple[str]
+    OUTPUT_TOOLTIPS: tuple[str, ...]
     """A tuple of strings to use as tooltips for node outputs, one for each item in `RETURN_TYPES`."""
     FUNCTION: str
     """The name of the function to execute as a literal string, e.g. `FUNCTION = "execute"`

comfy/k_diffusion/sampling.py

@@ -1345,28 +1345,52 @@ def sample_res_multistep_ancestral_cfg_pp(model, x, sigmas, extra_args=None, cal
     return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_noise=s_noise, noise_sampler=noise_sampler, eta=eta, cfg_pp=True)
 
 @torch.no_grad()
-def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2.):
+def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2., cfg_pp=False):
     """Gradient-estimation sampler. Paper: https://openreview.net/pdf?id=o2ND9v0CeK"""
     extra_args = {} if extra_args is None else extra_args
     s_in = x.new_ones([x.shape[0]])
     old_d = None
 
+    uncond_denoised = None
+    def post_cfg_function(args):
+        nonlocal uncond_denoised
+        uncond_denoised = args["uncond_denoised"]
+        return args["denoised"]
+
+    if cfg_pp:
+        model_options = extra_args.get("model_options", {}).copy()
+        extra_args["model_options"] = comfy.model_patcher.set_model_options_post_cfg_function(model_options, post_cfg_function, disable_cfg1_optimization=True)
+
     for i in trange(len(sigmas) - 1, disable=disable):
         denoised = model(x, sigmas[i] * s_in, **extra_args)
-        d = to_d(x, sigmas[i], denoised)
+        if cfg_pp:
+            d = to_d(x, sigmas[i], uncond_denoised)
+        else:
+            d = to_d(x, sigmas[i], denoised)
         if callback is not None:
             callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
         dt = sigmas[i + 1] - sigmas[i]
         if i == 0:
             # Euler method
-            x = x + d * dt
+            if cfg_pp:
+                x = denoised + d * sigmas[i + 1]
+            else:
+                x = x + d * dt
         else:
             # Gradient estimation
-            d_bar = ge_gamma * d + (1 - ge_gamma) * old_d
+            if cfg_pp:
-            x = x + d_bar * dt
+                d_bar = (ge_gamma - 1) * (d - old_d)
+                x = denoised + d * sigmas[i + 1] + d_bar * dt
+            else:
+                d_bar = ge_gamma * d + (1 - ge_gamma) * old_d
+                x = x + d_bar * dt
         old_d = d
     return x
 
+@torch.no_grad()
+def sample_gradient_estimation_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2.):
+    return sample_gradient_estimation(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, ge_gamma=ge_gamma, cfg_pp=True)
+
 @torch.no_grad()
 def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, s_noise=1., noise_sampler=None, noise_scaler=None, max_stage=3):
     """

comfy/ldm/chroma/layers.py

@@ -0,0 +1,183 @@
+import torch
+from torch import Tensor, nn
+
+from comfy.ldm.flux.math import attention
+from comfy.ldm.flux.layers import (
+    MLPEmbedder,
+    RMSNorm,
+    QKNorm,
+    SelfAttention,
+    ModulationOut,
+)
+
+
+
+class ChromaModulationOut(ModulationOut):
+    @classmethod
+    def from_offset(cls, tensor: torch.Tensor, offset: int = 0) -> ModulationOut:
+        return cls(
+            shift=tensor[:, offset : offset + 1, :],
+            scale=tensor[:, offset + 1 : offset + 2, :],
+            gate=tensor[:, offset + 2 : offset + 3, :],
+        )
+
+
+
+
+class Approximator(nn.Module):
+    def __init__(self, in_dim: int, out_dim: int, hidden_dim: int, n_layers = 5, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.in_proj = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device)
+        self.layers = nn.ModuleList([MLPEmbedder(hidden_dim, hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)])
+        self.norms = nn.ModuleList([RMSNorm(hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)])
+        self.out_proj = operations.Linear(hidden_dim, out_dim, dtype=dtype, device=device)
+
+    @property
+    def device(self):
+        # Get the device of the module (assumes all parameters are on the same device)
+        return next(self.parameters()).device
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = self.in_proj(x)
+
+        for layer, norms in zip(self.layers, self.norms):
+            x = x + layer(norms(x))
+
+        x = self.out_proj(x)
+
+        return x
+
+
+class DoubleStreamBlock(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, dtype=None, device=None, operations=None):
+        super().__init__()
+
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.img_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
+
+        self.img_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.img_mlp = nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+            nn.GELU(approximate="tanh"),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+
+        self.txt_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
+
+        self.txt_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.txt_mlp = nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+            nn.GELU(approximate="tanh"),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+        self.flipped_img_txt = flipped_img_txt
+
+    def forward(self, img: Tensor, txt: Tensor, pe: Tensor, vec: Tensor, attn_mask=None):
+        (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec
+
+        # prepare image for attention
+        img_modulated = self.img_norm1(img)
+        img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
+        img_qkv = self.img_attn.qkv(img_modulated)
+        img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
+
+        # prepare txt for attention
+        txt_modulated = self.txt_norm1(txt)
+        txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
+        txt_qkv = self.txt_attn.qkv(txt_modulated)
+        txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
+
+        # run actual attention
+        attn = attention(torch.cat((txt_q, img_q), dim=2),
+                         torch.cat((txt_k, img_k), dim=2),
+                         torch.cat((txt_v, img_v), dim=2),
+                         pe=pe, mask=attn_mask)
+
+        txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
+
+        # calculate the img bloks
+        img = img + img_mod1.gate * self.img_attn.proj(img_attn)
+        img = img + img_mod2.gate * self.img_mlp((1 + img_mod2.scale) * self.img_norm2(img) + img_mod2.shift)
+
+        # calculate the txt bloks
+        txt += txt_mod1.gate * self.txt_attn.proj(txt_attn)
+        txt += txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift)
+
+        if txt.dtype == torch.float16:
+            txt = torch.nan_to_num(txt, nan=0.0, posinf=65504, neginf=-65504)
+
+        return img, txt
+
+
+class SingleStreamBlock(nn.Module):
+    """
+    A DiT block with parallel linear layers as described in
+    https://arxiv.org/abs/2302.05442 and adapted modulation interface.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qk_scale: float = None,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.hidden_dim = hidden_size
+        self.num_heads = num_heads
+        head_dim = hidden_size // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        # qkv and mlp_in
+        self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim, dtype=dtype, device=device)
+        # proj and mlp_out
+        self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, dtype=dtype, device=device)
+
+        self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
+
+        self.hidden_size = hidden_size
+        self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+
+        self.mlp_act = nn.GELU(approximate="tanh")
+
+    def forward(self, x: Tensor, pe: Tensor, vec: Tensor, attn_mask=None) -> Tensor:
+        mod = vec
+        x_mod = (1 + mod.scale) * self.pre_norm(x) + mod.shift
+        qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
+
+        q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        q, k = self.norm(q, k, v)
+
+        # compute attention
+        attn = attention(q, k, v, pe=pe, mask=attn_mask)
+        # compute activation in mlp stream, cat again and run second linear layer
+        output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
+        x += mod.gate * output
+        if x.dtype == torch.float16:
+            x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
+        return x
+
+
+class LastLayer(nn.Module):
+    def __init__(self, hidden_size: int, patch_size: int, out_channels: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm_final = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.linear = operations.Linear(hidden_size, out_channels, bias=True, dtype=dtype, device=device)
+
+    def forward(self, x: Tensor, vec: Tensor) -> Tensor:
+        shift, scale = vec
+        shift = shift.squeeze(1)
+        scale = scale.squeeze(1)
+        x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
+        x = self.linear(x)
+        return x

comfy/ldm/chroma/model.py

@@ -0,0 +1,271 @@
+#Original code can be found on: https://github.com/black-forest-labs/flux
+
+from dataclasses import dataclass
+
+import torch
+from torch import Tensor, nn
+from einops import rearrange, repeat
+import comfy.ldm.common_dit
+
+from comfy.ldm.flux.layers import (
+    EmbedND,
+    timestep_embedding,
+)
+
+from .layers import (
+    DoubleStreamBlock,
+    LastLayer,
+    SingleStreamBlock,
+    Approximator,
+    ChromaModulationOut,
+)
+
+
+@dataclass
+class ChromaParams:
+    in_channels: int
+    out_channels: int
+    context_in_dim: int
+    hidden_size: int
+    mlp_ratio: float
+    num_heads: int
+    depth: int
+    depth_single_blocks: int
+    axes_dim: list
+    theta: int
+    patch_size: int
+    qkv_bias: bool
+    in_dim: int
+    out_dim: int
+    hidden_dim: int
+    n_layers: int
+
+
+
+
+class Chroma(nn.Module):
+    """
+    Transformer model for flow matching on sequences.
+    """
+
+    def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs):
+        super().__init__()
+        self.dtype = dtype
+        params = ChromaParams(**kwargs)
+        self.params = params
+        self.patch_size = params.patch_size
+        self.in_channels = params.in_channels
+        self.out_channels = params.out_channels
+        if params.hidden_size % params.num_heads != 0:
+            raise ValueError(
+                f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}"
+            )
+        pe_dim = params.hidden_size // params.num_heads
+        if sum(params.axes_dim) != pe_dim:
+            raise ValueError(f"Got {params.axes_dim} but expected positional dim {pe_dim}")
+        self.hidden_size = params.hidden_size
+        self.num_heads = params.num_heads
+        self.in_dim = params.in_dim
+        self.out_dim = params.out_dim
+        self.hidden_dim = params.hidden_dim
+        self.n_layers = params.n_layers
+        self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)
+        self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
+        self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device)
+        # set as nn identity for now, will overwrite it later.
+        self.distilled_guidance_layer = Approximator(
+                    in_dim=self.in_dim,
+                    hidden_dim=self.hidden_dim,
+                    out_dim=self.out_dim,
+                    n_layers=self.n_layers,
+                    dtype=dtype, device=device, operations=operations
+                )
+
+
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    self.hidden_size,
+                    self.num_heads,
+                    mlp_ratio=params.mlp_ratio,
+                    qkv_bias=params.qkv_bias,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(params.depth)
+            ]
+        )
+
+        self.single_blocks = nn.ModuleList(
+            [
+                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations)
+                for _ in range(params.depth_single_blocks)
+            ]
+        )
+
+        if final_layer:
+            self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, dtype=dtype, device=device, operations=operations)
+
+        self.skip_mmdit = []
+        self.skip_dit = []
+        self.lite = False
+
+    def get_modulations(self, tensor: torch.Tensor, block_type: str, *, idx: int = 0):
+        # This function slices up the modulations tensor which has the following layout:
+        #   single     : num_single_blocks * 3 elements
+        #   double_img : num_double_blocks * 6 elements
+        #   double_txt : num_double_blocks * 6 elements
+        #   final      : 2 elements
+        if block_type == "final":
+            return (tensor[:, -2:-1, :], tensor[:, -1:, :])
+        single_block_count = self.params.depth_single_blocks
+        double_block_count = self.params.depth
+        offset = 3 * idx
+        if block_type == "single":
+            return ChromaModulationOut.from_offset(tensor, offset)
+        # Double block modulations are 6 elements so we double 3 * idx.
+        offset *= 2
+        if block_type in {"double_img", "double_txt"}:
+            # Advance past the single block modulations.
+            offset += 3 * single_block_count
+            if block_type == "double_txt":
+                # Advance past the double block img modulations.
+                offset += 6 * double_block_count
+            return (
+                ChromaModulationOut.from_offset(tensor, offset),
+                ChromaModulationOut.from_offset(tensor, offset + 3),
+            )
+        raise ValueError("Bad block_type")
+
+
+    def forward_orig(
+        self,
+        img: Tensor,
+        img_ids: Tensor,
+        txt: Tensor,
+        txt_ids: Tensor,
+        timesteps: Tensor,
+        guidance: Tensor = None,
+        control = None,
+        transformer_options={},
+        attn_mask: Tensor = None,
+    ) -> Tensor:
+        patches_replace = transformer_options.get("patches_replace", {})
+        if img.ndim != 3 or txt.ndim != 3:
+            raise ValueError("Input img and txt tensors must have 3 dimensions.")
+
+        # running on sequences img
+        img = self.img_in(img)
+
+        # distilled vector guidance
+        mod_index_length = 344
+        distill_timestep = timestep_embedding(timesteps.detach().clone(), 16).to(img.device, img.dtype)
+        # guidance = guidance *
+        distil_guidance = timestep_embedding(guidance.detach().clone(), 16).to(img.device, img.dtype)
+
+        # get all modulation index
+        modulation_index = timestep_embedding(torch.arange(mod_index_length), 32).to(img.device, img.dtype)
+        # we need to broadcast the modulation index here so each batch has all of the index
+        modulation_index = modulation_index.unsqueeze(0).repeat(img.shape[0], 1, 1).to(img.device, img.dtype)
+        # and we need to broadcast timestep and guidance along too
+        timestep_guidance = torch.cat([distill_timestep, distil_guidance], dim=1).unsqueeze(1).repeat(1, mod_index_length, 1).to(img.dtype).to(img.device, img.dtype)
+        # then and only then we could concatenate it together
+        input_vec = torch.cat([timestep_guidance, modulation_index], dim=-1).to(img.device, img.dtype)
+
+        mod_vectors = self.distilled_guidance_layer(input_vec)
+
+        txt = self.txt_in(txt)
+
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        pe = self.pe_embedder(ids)
+
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.double_blocks):
+            if i not in self.skip_mmdit:
+                double_mod = (
+                    self.get_modulations(mod_vectors, "double_img", idx=i),
+                    self.get_modulations(mod_vectors, "double_txt", idx=i),
+                )
+                if ("double_block", i) in blocks_replace:
+                    def block_wrap(args):
+                        out = {}
+                        out["img"], out["txt"] = block(img=args["img"],
+                                                       txt=args["txt"],
+                                                       vec=args["vec"],
+                                                       pe=args["pe"],
+                                                       attn_mask=args.get("attn_mask"))
+                        return out
+
+                    out = blocks_replace[("double_block", i)]({"img": img,
+                                                               "txt": txt,
+                                                               "vec": double_mod,
+                                                               "pe": pe,
+                                                               "attn_mask": attn_mask},
+                                                              {"original_block": block_wrap})
+                    txt = out["txt"]
+                    img = out["img"]
+                else:
+                    img, txt = block(img=img,
+                                     txt=txt,
+                                     vec=double_mod,
+                                     pe=pe,
+                                     attn_mask=attn_mask)
+
+                if control is not None: # Controlnet
+                    control_i = control.get("input")
+                    if i < len(control_i):
+                        add = control_i[i]
+                        if add is not None:
+                            img += add
+
+        img = torch.cat((txt, img), 1)
+
+        for i, block in enumerate(self.single_blocks):
+            if i not in self.skip_dit:
+                single_mod = self.get_modulations(mod_vectors, "single", idx=i)
+                if ("single_block", i) in blocks_replace:
+                    def block_wrap(args):
+                        out = {}
+                        out["img"] = block(args["img"],
+                                           vec=args["vec"],
+                                           pe=args["pe"],
+                                           attn_mask=args.get("attn_mask"))
+                        return out
+
+                    out = blocks_replace[("single_block", i)]({"img": img,
+                                                               "vec": single_mod,
+                                                               "pe": pe,
+                                                               "attn_mask": attn_mask},
+                                                              {"original_block": block_wrap})
+                    img = out["img"]
+                else:
+                    img = block(img, vec=single_mod, pe=pe, attn_mask=attn_mask)
+
+                if control is not None: # Controlnet
+                    control_o = control.get("output")
+                    if i < len(control_o):
+                        add = control_o[i]
+                        if add is not None:
+                            img[:, txt.shape[1] :, ...] += add
+
+        img = img[:, txt.shape[1] :, ...]
+        final_mod = self.get_modulations(mod_vectors, "final")
+        img = self.final_layer(img, vec=final_mod)  # (N, T, patch_size ** 2 * out_channels)
+        return img
+
+    def forward(self, x, timestep, context, guidance, control=None, transformer_options={}, **kwargs):
+        bs, c, h, w = x.shape
+        patch_size = 2
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
+
+        img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
+
+        h_len = ((h + (patch_size // 2)) // patch_size)
+        w_len = ((w + (patch_size // 2)) // patch_size)
+        img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
+        img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
+
+        txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
+        out = self.forward_orig(img, img_ids, context, txt_ids, timestep, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None))
+        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)[:,:,:h,:w]

comfy/ldm/hidream/model.py

@@ -699,10 +699,13 @@ class HiDreamImageTransformer2DModel(nn.Module):
         y: Optional[torch.Tensor] = None,
         context: Optional[torch.Tensor] = None,
         encoder_hidden_states_llama3=None,
+        image_cond=None,
         control = None,
         transformer_options = {},
     ) -> torch.Tensor:
         bs, c, h, w = x.shape
+        if image_cond is not None:
+            x = torch.cat([x, image_cond], dim=-1)
         hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
         timesteps = t
         pooled_embeds = y

comfy/ldm/lightricks/model.py

@@ -1,7 +1,6 @@
 import torch
 from torch import nn
 import comfy.ldm.modules.attention
-from comfy.ldm.genmo.joint_model.layers import RMSNorm
 import comfy.ldm.common_dit
 from einops import rearrange
 import math
@@ -262,8 +261,8 @@ class CrossAttention(nn.Module):
         self.heads = heads
         self.dim_head = dim_head
 
-        self.q_norm = RMSNorm(inner_dim, dtype=dtype, device=device)
+        self.q_norm = operations.RMSNorm(inner_dim, dtype=dtype, device=device)
-        self.k_norm = RMSNorm(inner_dim, dtype=dtype, device=device)
+        self.k_norm = operations.RMSNorm(inner_dim, dtype=dtype, device=device)
 
         self.to_q = operations.Linear(query_dim, inner_dim, bias=True, dtype=dtype, device=device)
         self.to_k = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device)

comfy/ldm/wan/model.py

@@ -631,6 +631,7 @@ class VaceWanModel(WanModel):
             if ii is not None:
                 c_skip, c = self.vace_blocks[ii](c, x=x_orig, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
                 x += c_skip * vace_strength
+                del c_skip
         # head
         x = self.head(x, e)
 

comfy/model_base.py

@@ -38,6 +38,7 @@ import comfy.ldm.lumina.model
 import comfy.ldm.wan.model
 import comfy.ldm.hunyuan3d.model
 import comfy.ldm.hidream.model
+import comfy.ldm.chroma.model
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -786,8 +787,8 @@ class PixArt(BaseModel):
         return out
 
 class Flux(BaseModel):
-    def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
+    def __init__(self, model_config, model_type=ModelType.FLUX, device=None, unet_model=comfy.ldm.flux.model.Flux):
-        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.flux.model.Flux)
+        super().__init__(model_config, model_type, device=device, unet_model=unet_model)
 
     def concat_cond(self, **kwargs):
         try:
@@ -1104,4 +1105,19 @@ class HiDream(BaseModel):
         conditioning_llama3 = kwargs.get("conditioning_llama3", None)
         if conditioning_llama3 is not None:
             out['encoder_hidden_states_llama3'] = comfy.conds.CONDRegular(conditioning_llama3)
+        image_cond = kwargs.get("concat_latent_image", None)
+        if image_cond is not None:
+            out['image_cond'] = comfy.conds.CONDNoiseShape(self.process_latent_in(image_cond))
+        return out
+
+class Chroma(Flux):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.chroma.model.Chroma)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+
+        guidance = kwargs.get("guidance", 0)
+        if guidance is not None:
+            out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
         return out

comfy/model_detection.py

@@ -164,7 +164,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         if in_key in state_dict_keys:
             dit_config["in_channels"] = state_dict[in_key].shape[1] // (patch_size * patch_size)
         dit_config["out_channels"] = 16
-        dit_config["vec_in_dim"] = 768
+        vec_in_key = '{}vector_in.in_layer.weight'.format(key_prefix)
+        if vec_in_key in state_dict_keys:
+            dit_config["vec_in_dim"] = state_dict[vec_in_key].shape[1]
         dit_config["context_in_dim"] = 4096
         dit_config["hidden_size"] = 3072
         dit_config["mlp_ratio"] = 4.0
@@ -174,7 +176,16 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["axes_dim"] = [16, 56, 56]
         dit_config["theta"] = 10000
         dit_config["qkv_bias"] = True
-        dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
+        if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma
+            dit_config["image_model"] = "chroma"
+            dit_config["in_channels"] = 64
+            dit_config["out_channels"] = 64
+            dit_config["in_dim"] = 64
+            dit_config["out_dim"] = 3072
+            dit_config["hidden_dim"] = 5120
+            dit_config["n_layers"] = 5
+        else:
+            dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
         return dit_config
 
     if '{}t5_yproj.weight'.format(key_prefix) in state_dict_keys: #Genmo mochi preview

comfy/model_management.py

@@ -946,9 +946,9 @@ if args.async_offload:
     NUM_STREAMS = 2
     logging.info("Using async weight offloading with {} streams".format(NUM_STREAMS))
 
-stream_counter = 0
+stream_counters = {}
 def get_offload_stream(device):
-    global stream_counter
+    stream_counter = stream_counters.get(device, 0)
     if NUM_STREAMS <= 1:
         return None
 
@@ -958,14 +958,16 @@ def get_offload_stream(device):
         stream_counter = (stream_counter + 1) % len(ss)
         if is_device_cuda(device):
             ss[stream_counter].wait_stream(torch.cuda.current_stream())
+        stream_counters[device] = stream_counter
         return s
     elif is_device_cuda(device):
         ss = []
         for k in range(NUM_STREAMS):
-            ss.append(torch.cuda.Stream(device=device, priority=10))
+            ss.append(torch.cuda.Stream(device=device, priority=0))
         STREAMS[device] = ss
         s = ss[stream_counter]
         stream_counter = (stream_counter + 1) % len(ss)
+        stream_counters[device] = stream_counter
         return s
     return None
 

comfy/model_sampling.py

@@ -111,13 +111,14 @@ class ModelSamplingDiscrete(torch.nn.Module):
         self.num_timesteps = int(timesteps)
         self.linear_start = linear_start
         self.linear_end = linear_end
+        self.zsnr = zsnr
 
         # self.register_buffer('betas', torch.tensor(betas, dtype=torch.float32))
         # self.register_buffer('alphas_cumprod', torch.tensor(alphas_cumprod, dtype=torch.float32))
         # self.register_buffer('alphas_cumprod_prev', torch.tensor(alphas_cumprod_prev, dtype=torch.float32))
 
         sigmas = ((1 - alphas_cumprod) / alphas_cumprod) ** 0.5
-        if zsnr:
+        if self.zsnr:
             sigmas = rescale_zero_terminal_snr_sigmas(sigmas)
 
         self.set_sigmas(sigmas)

comfy/samplers.py

@@ -710,7 +710,7 @@ KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_c
                   "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_2s_ancestral_cfg_pp", "dpmpp_sde", "dpmpp_sde_gpu",
                   "dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm",
                   "ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp", "res_multistep_ancestral", "res_multistep_ancestral_cfg_pp",
-                  "gradient_estimation", "er_sde", "seeds_2", "seeds_3"]
+                  "gradient_estimation", "gradient_estimation_cfg_pp", "er_sde", "seeds_2", "seeds_3"]
 
 class KSAMPLER(Sampler):
     def __init__(self, sampler_function, extra_options={}, inpaint_options={}):

comfy/sd.py

@@ -714,6 +714,7 @@ class CLIPType(Enum):
     LUMINA2 = 12
     WAN = 13
     HIDREAM = 14
+    CHROMA = 15
 
 
 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
@@ -818,7 +819,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             elif clip_type == CLIPType.LTXV:
                 clip_target.clip = comfy.text_encoders.lt.ltxv_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.lt.LTXVT5Tokenizer
-            elif clip_type == CLIPType.PIXART:
+            elif clip_type == CLIPType.PIXART or clip_type == CLIPType.CHROMA:
                 clip_target.clip = comfy.text_encoders.pixart_t5.pixart_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.pixart_t5.PixArtTokenizer
             elif clip_type == CLIPType.WAN:

comfy/supported_models.py

@@ -993,6 +993,10 @@ class WAN21_Vace(WAN21_T2V):
         "model_type": "vace",
     }
 
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+        self.memory_usage_factor = 1.2 * self.memory_usage_factor
+
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.WAN21_Vace(self, image_to_video=False, device=device)
         return out
@@ -1064,7 +1068,34 @@ class HiDream(supported_models_base.BASE):
     def clip_target(self, state_dict={}):
         return None #  TODO
 
+class Chroma(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "chroma",
+    }
 
-models = [LotusD, 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, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream]
+    unet_extra_config = {
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+    }
+
+    latent_format = comfy.latent_formats.Flux
+
+    memory_usage_factor = 3.2
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.Chroma(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.pixart_t5.PixArtTokenizer, comfy.text_encoders.pixart_t5.pixart_te(**t5_detect))
+
+models = [LotusD, 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, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma]
 
 models += [SVD_img2vid]

comfy_api/input/__init__.py

@@ -0,0 +1,8 @@
+from .basic_types import ImageInput, AudioInput
+from .video_types import VideoInput
+
+__all__ = [
+    "ImageInput",
+    "AudioInput",
+    "VideoInput",
+]

comfy_api/input/basic_types.py

@@ -0,0 +1,20 @@
+import torch
+from typing import TypedDict
+
+ImageInput = torch.Tensor
+"""
+An image in format [B, H, W, C] where B is the batch size, C is the number of channels,
+"""
+
+class AudioInput(TypedDict):
+    """
+    TypedDict representing audio input.
+    """
+
+    waveform: torch.Tensor
+    """
+    Tensor in the format [B, C, T] where B is the batch size, C is the number of channels,
+    """
+
+    sample_rate: int
+

comfy_api/input/video_types.py

@@ -0,0 +1,45 @@
+from __future__ import annotations
+from abc import ABC, abstractmethod
+from typing import Optional
+from comfy_api.util import VideoContainer, VideoCodec, VideoComponents
+
+class VideoInput(ABC):
+    """
+    Abstract base class for video input types.
+    """
+
+    @abstractmethod
+    def get_components(self) -> VideoComponents:
+        """
+        Abstract method to get the video components (images, audio, and frame rate).
+
+        Returns:
+            VideoComponents containing images, audio, and frame rate
+        """
+        pass
+
+    @abstractmethod
+    def save_to(
+        self,
+        path: str,
+        format: VideoContainer = VideoContainer.AUTO,
+        codec: VideoCodec = VideoCodec.AUTO,
+        metadata: Optional[dict] = None
+    ):
+        """
+        Abstract method to save the video input to a file.
+        """
+        pass
+
+    # Provide a default implementation, but subclasses can provide optimized versions
+    # if possible.
+    def get_dimensions(self) -> tuple[int, int]:
+        """
+        Returns the dimensions of the video input.
+
+        Returns:
+            Tuple of (width, height)
+        """
+        components = self.get_components()
+        return components.images.shape[2], components.images.shape[1]
+

comfy_api/input_impl/__init__.py

@@ -0,0 +1,7 @@
+from .video_types import VideoFromFile, VideoFromComponents
+
+__all__ = [
+    # Implementations
+    "VideoFromFile",
+    "VideoFromComponents",
+]

comfy_api/input_impl/video_types.py

@@ -0,0 +1,224 @@
+from __future__ import annotations
+from av.container import InputContainer
+from av.subtitles.stream import SubtitleStream
+from fractions import Fraction
+from typing import Optional
+from comfy_api.input import AudioInput
+import av
+import io
+import json
+import numpy as np
+import torch
+from comfy_api.input import VideoInput
+from comfy_api.util import VideoContainer, VideoCodec, VideoComponents
+
+class VideoFromFile(VideoInput):
+    """
+    Class representing video input from a file.
+    """
+
+    def __init__(self, file: str | io.BytesIO):
+        """
+        Initialize the VideoFromFile object based off of either a path on disk or a BytesIO object
+        containing the file contents.
+        """
+        self.__file = file
+
+    def get_dimensions(self) -> tuple[int, int]:
+        """
+        Returns the dimensions of the video input.
+
+        Returns:
+            Tuple of (width, height)
+        """
+        if isinstance(self.__file, io.BytesIO):
+            self.__file.seek(0)  # Reset the BytesIO object to the beginning
+        with av.open(self.__file, mode='r') as container:
+            for stream in container.streams:
+                if stream.type == 'video':
+                    assert isinstance(stream, av.VideoStream)
+                    return stream.width, stream.height
+        raise ValueError(f"No video stream found in file '{self.__file}'")
+
+    def get_components_internal(self, container: InputContainer) -> VideoComponents:
+        # Get video frames
+        frames = []
+        for frame in container.decode(video=0):
+            img = frame.to_ndarray(format='rgb24')  # shape: (H, W, 3)
+            img = torch.from_numpy(img) / 255.0  # shape: (H, W, 3)
+            frames.append(img)
+
+        images = torch.stack(frames) if len(frames) > 0 else torch.zeros(0, 3, 0, 0)
+
+        # Get frame rate
+        video_stream = next(s for s in container.streams if s.type == 'video')
+        frame_rate = Fraction(video_stream.average_rate) if video_stream and video_stream.average_rate else Fraction(1)
+
+        # Get audio if available
+        audio = None
+        try:
+            container.seek(0)  # Reset the container to the beginning
+            for stream in container.streams:
+                if stream.type != 'audio':
+                    continue
+                assert isinstance(stream, av.AudioStream)
+                audio_frames = []
+                for packet in container.demux(stream):
+                    for frame in packet.decode():
+                        assert isinstance(frame, av.AudioFrame)
+                        audio_frames.append(frame.to_ndarray())  # shape: (channels, samples)
+                if len(audio_frames) > 0:
+                    audio_data = np.concatenate(audio_frames, axis=1)  # shape: (channels, total_samples)
+                    audio_tensor = torch.from_numpy(audio_data).unsqueeze(0)  # shape: (1, channels, total_samples)
+                    audio = AudioInput({
+                        "waveform": audio_tensor,
+                        "sample_rate": int(stream.sample_rate) if stream.sample_rate else 1,
+                    })
+        except StopIteration:
+            pass  # No audio stream
+
+        metadata = container.metadata
+        return VideoComponents(images=images, audio=audio, frame_rate=frame_rate, metadata=metadata)
+
+    def get_components(self) -> VideoComponents:
+        if isinstance(self.__file, io.BytesIO):
+            self.__file.seek(0)  # Reset the BytesIO object to the beginning
+        with av.open(self.__file, mode='r') as container:
+            return self.get_components_internal(container)
+        raise ValueError(f"No video stream found in file '{self.__file}'")
+
+    def save_to(
+        self,
+        path: str,
+        format: VideoContainer = VideoContainer.AUTO,
+        codec: VideoCodec = VideoCodec.AUTO,
+        metadata: Optional[dict] = None
+    ):
+        if isinstance(self.__file, io.BytesIO):
+            self.__file.seek(0)  # Reset the BytesIO object to the beginning
+        with av.open(self.__file, mode='r') as container:
+            container_format = container.format.name
+            video_encoding = container.streams.video[0].codec.name if len(container.streams.video) > 0 else None
+            reuse_streams = True
+            if format != VideoContainer.AUTO and format not in container_format.split(","):
+                reuse_streams = False
+            if codec != VideoCodec.AUTO and codec != video_encoding and video_encoding is not None:
+                reuse_streams = False
+
+            if not reuse_streams:
+                components = self.get_components_internal(container)
+                video = VideoFromComponents(components)
+                return video.save_to(
+                    path,
+                    format=format,
+                    codec=codec,
+                    metadata=metadata
+                )
+
+            streams = container.streams
+            with av.open(path, mode='w', options={"movflags": "use_metadata_tags"}) as output_container:
+                # Copy over the original metadata
+                for key, value in container.metadata.items():
+                    if metadata is None or key not in metadata:
+                        output_container.metadata[key] = value
+
+                # Add our new metadata
+                if metadata is not None:
+                    for key, value in metadata.items():
+                        if isinstance(value, str):
+                            output_container.metadata[key] = value
+                        else:
+                            output_container.metadata[key] = json.dumps(value)
+
+                # Add streams to the new container
+                stream_map = {}
+                for stream in streams:
+                    if isinstance(stream, (av.VideoStream, av.AudioStream, SubtitleStream)):
+                        out_stream = output_container.add_stream_from_template(template=stream, opaque=True)
+                        stream_map[stream] = out_stream
+
+                # Write packets to the new container
+                for packet in container.demux():
+                    if packet.stream in stream_map and packet.dts is not None:
+                        packet.stream = stream_map[packet.stream]
+                        output_container.mux(packet)
+
+class VideoFromComponents(VideoInput):
+    """
+    Class representing video input from tensors.
+    """
+
+    def __init__(self, components: VideoComponents):
+        self.__components = components
+
+    def get_components(self) -> VideoComponents:
+        return VideoComponents(
+            images=self.__components.images,
+            audio=self.__components.audio,
+            frame_rate=self.__components.frame_rate
+        )
+
+    def save_to(
+        self,
+        path: str,
+        format: VideoContainer = VideoContainer.AUTO,
+        codec: VideoCodec = VideoCodec.AUTO,
+        metadata: Optional[dict] = None
+    ):
+        if format != VideoContainer.AUTO and format != VideoContainer.MP4:
+            raise ValueError("Only MP4 format is supported for now")
+        if codec != VideoCodec.AUTO and codec != VideoCodec.H264:
+            raise ValueError("Only H264 codec is supported for now")
+        with av.open(path, mode='w', options={'movflags': 'use_metadata_tags'}) as output:
+            # Add metadata before writing any streams
+            if metadata is not None:
+                for key, value in metadata.items():
+                    output.metadata[key] = json.dumps(value)
+
+            frame_rate = Fraction(round(self.__components.frame_rate * 1000), 1000)
+            # Create a video stream
+            video_stream = output.add_stream('h264', rate=frame_rate)
+            video_stream.width = self.__components.images.shape[2]
+            video_stream.height = self.__components.images.shape[1]
+            video_stream.pix_fmt = 'yuv420p'
+
+            # Create an audio stream
+            audio_sample_rate = 1
+            audio_stream: Optional[av.AudioStream] = None
+            if self.__components.audio:
+                audio_sample_rate = int(self.__components.audio['sample_rate'])
+                audio_stream = output.add_stream('aac', rate=audio_sample_rate)
+                audio_stream.sample_rate = audio_sample_rate
+                audio_stream.format = 'fltp'
+
+            # Encode video
+            for i, frame in enumerate(self.__components.images):
+                img = (frame * 255).clamp(0, 255).byte().cpu().numpy() # shape: (H, W, 3)
+                frame = av.VideoFrame.from_ndarray(img, format='rgb24')
+                frame = frame.reformat(format='yuv420p')  # Convert to YUV420P as required by h264
+                packet = video_stream.encode(frame)
+                output.mux(packet)
+
+            # Flush video
+            packet = video_stream.encode(None)
+            output.mux(packet)
+
+            if audio_stream and self.__components.audio:
+                # Encode audio
+                samples_per_frame = int(audio_sample_rate / frame_rate)
+                num_frames = self.__components.audio['waveform'].shape[2] // samples_per_frame
+                for i in range(num_frames):
+                    start = i * samples_per_frame
+                    end = start + samples_per_frame
+                    # TODO(Feature) - Add support for stereo audio
+                    chunk = self.__components.audio['waveform'][0, 0, start:end].unsqueeze(0).numpy()
+                    audio_frame = av.AudioFrame.from_ndarray(chunk, format='fltp', layout='mono')
+                    audio_frame.sample_rate = audio_sample_rate
+                    audio_frame.pts = i * samples_per_frame
+                    for packet in audio_stream.encode(audio_frame):
+                        output.mux(packet)
+
+                # Flush audio
+                for packet in audio_stream.encode(None):
+                    output.mux(packet)
+

comfy_api/util/__init__.py

@@ -0,0 +1,8 @@
+from .video_types import VideoContainer, VideoCodec, VideoComponents
+
+__all__ = [
+    # Utility Types
+    "VideoContainer",
+    "VideoCodec",
+    "VideoComponents",
+]

comfy_api/util/video_types.py

@@ -0,0 +1,51 @@
+from __future__ import annotations
+from dataclasses import dataclass
+from enum import Enum
+from fractions import Fraction
+from typing import Optional
+from comfy_api.input import ImageInput, AudioInput
+
+class VideoCodec(str, Enum):
+    AUTO = "auto"
+    H264 = "h264"
+
+    @classmethod
+    def as_input(cls) -> list[str]:
+        """
+        Returns a list of codec names that can be used as node input.
+        """
+        return [member.value for member in cls]
+
+class VideoContainer(str, Enum):
+    AUTO = "auto"
+    MP4 = "mp4"
+
+    @classmethod
+    def as_input(cls) -> list[str]:
+        """
+        Returns a list of container names that can be used as node input.
+        """
+        return [member.value for member in cls]
+
+    @classmethod
+    def get_extension(cls, value) -> str:
+        """
+        Returns the file extension for the container.
+        """
+        if isinstance(value, str):
+            value = cls(value)
+        if value == VideoContainer.MP4 or value == VideoContainer.AUTO:
+            return "mp4"
+        return ""
+
+@dataclass
+class VideoComponents:
+    """
+    Dataclass representing the components of a video.
+    """
+
+    images: ImageInput
+    frame_rate: Fraction
+    audio: Optional[AudioInput] = None
+    metadata: Optional[dict] = None
+

comfy_api_nodes/apis/client.py

@@ -297,6 +297,10 @@ class SynchronousOperation(Generic[T, R]):
 
             # Convert request model to dict, but use None for EmptyRequest
             request_dict = None if isinstance(self.request, EmptyRequest) else self.request.model_dump(exclude_none=True)
+            if request_dict:
+                for key, value in request_dict.items():
+                    if isinstance(value, Enum):
+                        request_dict[key] = value.value
 
             # Debug log for request
             logging.debug(f"[DEBUG] API Request: {self.endpoint.method.value} {self.endpoint.path}")

comfy_api_nodes/nodes_api.py

@@ -1,21 +1,22 @@
+import base64
 import io
+import math
 from inspect import cleandoc
 
-from comfy.utils import common_upscale
+import numpy as np
+import requests
+import torch
+from PIL import Image
+
 from comfy.comfy_types.node_typing import IO, ComfyNodeABC, InputTypeDict
+from comfy.utils import common_upscale
 from comfy_api_nodes.apis import (
-    OpenAIImageGenerationRequest,
     OpenAIImageEditRequest,
-    OpenAIImageGenerationResponse
+    OpenAIImageGenerationRequest,
+    OpenAIImageGenerationResponse,
 )
 from comfy_api_nodes.apis.client import ApiEndpoint, HttpMethod, SynchronousOperation
 
-import numpy as np
-from PIL import Image
-import requests
-import torch
-import math
-import base64
 
 def downscale_input(image):
     samples = image.movedim(-1,1)
@@ -331,6 +332,11 @@ class OpenAIGPTImage1(ComfyNodeABC):
                     "default": None,
                     "tooltip": "Optional mask for inpainting (white areas will be replaced)",
                 }),
+                "moderation": (IO.COMBO, {
+                    "options": ["low","auto"],
+                    "default": "low",
+                    "tooltip": "Moderation level",
+                }),
             },
             "hidden": {
                 "auth_token": "AUTH_TOKEN_COMFY_ORG"
@@ -343,7 +349,7 @@ class OpenAIGPTImage1(ComfyNodeABC):
     DESCRIPTION = cleandoc(__doc__ or "")
     API_NODE = True
 
-    def api_call(self, prompt, seed=0, quality="low", background="opaque", image=None, mask=None, n=1, size="1024x1024", auth_token=None):
+    def api_call(self, prompt, seed=0, quality="low", background="opaque", image=None, mask=None, n=1, size="1024x1024", auth_token=None, moderation="low"):
         model = "gpt-image-1"
         path = "/proxy/openai/images/generations"
         request_class = OpenAIImageGenerationRequest
@@ -415,6 +421,7 @@ class OpenAIGPTImage1(ComfyNodeABC):
                 n=n,
                 seed=seed,
                 size=size,
+                moderation=moderation,
             ),
             files=files if files else None,
             auth_token=auth_token

comfy_extras/nodes_lt.py

@@ -38,6 +38,7 @@ class LTXVImgToVideo:
                              "height": ("INT", {"default": 512, "min": 64, "max": nodes.MAX_RESOLUTION, "step": 32}),
                              "length": ("INT", {"default": 97, "min": 9, "max": nodes.MAX_RESOLUTION, "step": 8}),
                              "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
+                             "strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0}),
                              }}
 
     RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
@@ -46,7 +47,7 @@ class LTXVImgToVideo:
     CATEGORY = "conditioning/video_models"
     FUNCTION = "generate"
 
-    def generate(self, positive, negative, image, vae, width, height, length, batch_size):
+    def generate(self, positive, negative, image, vae, width, height, length, batch_size, strength):
         pixels = comfy.utils.common_upscale(image.movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
         encode_pixels = pixels[:, :, :, :3]
         t = vae.encode(encode_pixels)
@@ -59,7 +60,7 @@ class LTXVImgToVideo:
             dtype=torch.float32,
             device=latent.device,
         )
-        conditioning_latent_frames_mask[:, :, :t.shape[2]] = 0
+        conditioning_latent_frames_mask[:, :, :t.shape[2]] = 1.0 - strength
 
         return (positive, negative, {"samples": latent, "noise_mask": conditioning_latent_frames_mask}, )
 
@@ -152,6 +153,15 @@ class LTXVAddGuide:
         return node_helpers.conditioning_set_values(cond, {"keyframe_idxs": keyframe_idxs})
 
     def append_keyframe(self, positive, negative, frame_idx, latent_image, noise_mask, guiding_latent, strength, scale_factors):
+        _, latent_idx = self.get_latent_index(
+            cond=positive,
+            latent_length=latent_image.shape[2],
+            guide_length=guiding_latent.shape[2],
+            frame_idx=frame_idx,
+            scale_factors=scale_factors,
+        )
+        noise_mask[:, :, latent_idx:latent_idx + guiding_latent.shape[2]] = 1.0
+
         positive = self.add_keyframe_index(positive, frame_idx, guiding_latent, scale_factors)
         negative = self.add_keyframe_index(negative, frame_idx, guiding_latent, scale_factors)
 

comfy_extras/nodes_model_merging.py

@@ -209,6 +209,9 @@ def save_checkpoint(model, clip=None, vae=None, clip_vision=None, filename_prefi
         metadata["modelspec.predict_key"] = "epsilon"
     elif model.model.model_type == comfy.model_base.ModelType.V_PREDICTION:
         metadata["modelspec.predict_key"] = "v"
+        extra_keys["v_pred"] = torch.tensor([])
+        if getattr(model_sampling, "zsnr", False):
+            extra_keys["ztsnr"] = torch.tensor([])
 
     if not args.disable_metadata:
         metadata["prompt"] = prompt_info

comfy_extras/nodes_optimalsteps.py

@@ -20,13 +20,14 @@ def loglinear_interp(t_steps, num_steps):
 
 NOISE_LEVELS = {"FLUX": [0.9968, 0.9886, 0.9819, 0.975, 0.966, 0.9471, 0.9158, 0.8287, 0.5512, 0.2808, 0.001],
 "Wan":[1.0, 0.997, 0.995, 0.993, 0.991, 0.989, 0.987, 0.985, 0.98, 0.975, 0.973, 0.968, 0.96, 0.946, 0.927, 0.902, 0.864, 0.776, 0.539, 0.208, 0.001],
+"Chroma": [0.992, 0.99, 0.988, 0.985, 0.982, 0.978, 0.973, 0.968, 0.961, 0.953, 0.943, 0.931, 0.917, 0.9, 0.881, 0.858, 0.832, 0.802, 0.769, 0.731, 0.69, 0.646, 0.599, 0.55, 0.501, 0.451, 0.402, 0.355, 0.311, 0.27, 0.232, 0.199, 0.169, 0.143, 0.12, 0.101, 0.084, 0.07, 0.058, 0.048, 0.001],
 }
 
 class OptimalStepsScheduler:
     @classmethod
     def INPUT_TYPES(s):
         return {"required":
-                    {"model_type": (["FLUX", "Wan"], ),
+                    {"model_type": (["FLUX", "Wan", "Chroma"], ),
                      "steps": ("INT", {"default": 20, "min": 3, "max": 1000}),
                      "denoise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}),
                       }

comfy_extras/nodes_post_processing.py

@@ -141,6 +141,7 @@ class Quantize:
 
     CATEGORY = "image/postprocessing"
 
+    @staticmethod
     def bayer(im, pal_im, order):
         def normalized_bayer_matrix(n):
             if n == 0:

comfy_extras/nodes_video.py

@@ -5,9 +5,13 @@ import av
 import torch
 import folder_paths
 import json
+from typing import Optional, Literal
 from fractions import Fraction
-from comfy.comfy_types import FileLocator
+from comfy.comfy_types import IO, FileLocator, ComfyNodeABC
-
+from comfy_api.input import ImageInput, AudioInput, VideoInput
+from comfy_api.util import VideoContainer, VideoCodec, VideoComponents
+from comfy_api.input_impl import VideoFromFile, VideoFromComponents
+from comfy.cli_args import args
 
 class SaveWEBM:
     def __init__(self):
@@ -75,7 +79,163 @@ class SaveWEBM:
 
         return {"ui": {"images": results, "animated": (True,)}}  # TODO: frontend side
 
+class SaveVideo(ComfyNodeABC):
+    def __init__(self):
+        self.output_dir = folder_paths.get_output_directory()
+        self.type: Literal["output"] = "output"
+        self.prefix_append = ""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "video": (IO.VIDEO, {"tooltip": "The video to save."}),
+                "filename_prefix": ("STRING", {"default": "video/ComfyUI", "tooltip": "The prefix for the file to save. This may include formatting information such as %date:yyyy-MM-dd% or %Empty Latent Image.width% to include values from nodes."}),
+                "format": (VideoContainer.as_input(), {"default": "auto", "tooltip": "The format to save the video as."}),
+                "codec": (VideoCodec.as_input(), {"default": "auto", "tooltip": "The codec to use for the video."}),
+            },
+            "hidden": {
+                "prompt": "PROMPT",
+                "extra_pnginfo": "EXTRA_PNGINFO"
+            },
+        }
+
+    RETURN_TYPES = ()
+    FUNCTION = "save_video"
+
+    OUTPUT_NODE = True
+
+    CATEGORY = "image/video"
+    DESCRIPTION = "Saves the input images to your ComfyUI output directory."
+
+    def save_video(self, video: VideoInput, filename_prefix, format, codec, prompt=None, extra_pnginfo=None):
+        filename_prefix += self.prefix_append
+        width, height = video.get_dimensions()
+        full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(
+            filename_prefix,
+            self.output_dir,
+            width,
+            height
+        )
+        results: list[FileLocator] = list()
+        saved_metadata = None
+        if not args.disable_metadata:
+            metadata = {}
+            if extra_pnginfo is not None:
+                metadata.update(extra_pnginfo)
+            if prompt is not None:
+                metadata["prompt"] = prompt
+            if len(metadata) > 0:
+                saved_metadata = metadata
+        file = f"{filename}_{counter:05}_.{VideoContainer.get_extension(format)}"
+        video.save_to(
+            os.path.join(full_output_folder, file),
+            format=format,
+            codec=codec,
+            metadata=saved_metadata
+        )
+
+        results.append({
+            "filename": file,
+            "subfolder": subfolder,
+            "type": self.type
+        })
+        counter += 1
+
+        return { "ui": { "images": results, "animated": (True,) } }
+
+class CreateVideo(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "images": (IO.IMAGE, {"tooltip": "The images to create a video from."}),
+                "fps": ("FLOAT", {"default": 30.0, "min": 1.0, "max": 120.0, "step": 1.0}),
+            },
+            "optional": {
+                "audio": (IO.AUDIO, {"tooltip": "The audio to add to the video."}),
+            }
+        }
+
+    RETURN_TYPES = (IO.VIDEO,)
+    FUNCTION = "create_video"
+
+    CATEGORY = "image/video"
+    DESCRIPTION = "Create a video from images."
+
+    def create_video(self, images: ImageInput, fps: float, audio: Optional[AudioInput] = None):
+        return (VideoFromComponents(
+            VideoComponents(
+            images=images,
+            audio=audio,
+            frame_rate=Fraction(fps),
+            )
+        ),)
+
+class GetVideoComponents(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "video": (IO.VIDEO, {"tooltip": "The video to extract components from."}),
+            }
+        }
+    RETURN_TYPES = (IO.IMAGE, IO.AUDIO, IO.FLOAT)
+    RETURN_NAMES = ("images", "audio", "fps")
+    FUNCTION = "get_components"
+
+    CATEGORY = "image/video"
+    DESCRIPTION = "Extracts all components from a video: frames, audio, and framerate."
+
+    def get_components(self, video: VideoInput):
+        components = video.get_components()
+
+        return (components.images, components.audio, float(components.frame_rate))
+
+class LoadVideo(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls):
+        input_dir = folder_paths.get_input_directory()
+        files = [f for f in os.listdir(input_dir) if os.path.isfile(os.path.join(input_dir, f))]
+        files = folder_paths.filter_files_content_types(files, ["video"])
+        return {"required":
+                    {"file": (sorted(files), {"video_upload": True})},
+                }
+
+    CATEGORY = "image/video"
+
+    RETURN_TYPES = (IO.VIDEO,)
+    FUNCTION = "load_video"
+    def load_video(self, file):
+        video_path = folder_paths.get_annotated_filepath(file)
+        return (VideoFromFile(video_path),)
+
+    @classmethod
+    def IS_CHANGED(cls, file):
+        video_path = folder_paths.get_annotated_filepath(file)
+        mod_time = os.path.getmtime(video_path)
+        # Instead of hashing the file, we can just use the modification time to avoid
+        # rehashing large files.
+        return mod_time
+
+    @classmethod
+    def VALIDATE_INPUTS(cls, file):
+        if not folder_paths.exists_annotated_filepath(file):
+            return "Invalid video file: {}".format(file)
+
+        return True
 
 NODE_CLASS_MAPPINGS = {
     "SaveWEBM": SaveWEBM,
+    "SaveVideo": SaveVideo,
+    "CreateVideo": CreateVideo,
+    "GetVideoComponents": GetVideoComponents,
+    "LoadVideo": LoadVideo,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "SaveVideo": "Save Video",
+    "CreateVideo": "Create Video",
+    "GetVideoComponents": "Get Video Components",
+    "LoadVideo": "Load Video",
 }

comfy_extras/nodes_webcam.py

@@ -20,7 +20,7 @@ class WebcamCapture(nodes.LoadImage):
 
     CATEGORY = "image"
 
-    def load_capture(s, image, **kwargs):
+    def load_capture(self, image, **kwargs):
         return super().load_image(folder_paths.get_annotated_filepath(image))
 
 

folder_paths.py

@@ -4,7 +4,7 @@ import os
 import time
 import mimetypes
 import logging
-from typing import Literal
+from typing import Literal, List
 from collections.abc import Collection
 
 from comfy.cli_args import args
@@ -141,7 +141,7 @@ def get_directory_by_type(type_name: str) -> str | None:
         return get_input_directory()
     return None
 
-def filter_files_content_types(files: list[str], content_types: Literal["image", "video", "audio", "model"]) -> list[str]:
+def filter_files_content_types(files: list[str], content_types: List[Literal["image", "video", "audio", "model"]]) -> list[str]:
     """
     Example:
         files = os.listdir(folder_paths.get_input_directory())

nodes.py

@@ -918,7 +918,7 @@ class CLIPLoader:
     @classmethod
     def INPUT_TYPES(s):
         return {"required": { "clip_name": (folder_paths.get_filename_list("text_encoders"), ),
-                              "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream"], ),
+                              "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma"], ),
                               },
                 "optional": {
                               "device": (["default", "cpu"], {"advanced": True}),

pyproject.toml

@@ -12,6 +12,7 @@ documentation = "https://docs.comfy.org/"
 
 [tool.ruff]
 lint.select = [
+  "N805", # invalid-first-argument-name-for-method
   "S307", # suspicious-eval-usage
   "S102", # exec
   "T",    # print-usage

requirements.txt

@@ -1,4 +1,4 @@
-comfyui-frontend-package==1.17.11
+comfyui-frontend-package==1.18.5
 comfyui-workflow-templates==0.1.3
 torch
 torchsde
@@ -22,5 +22,5 @@ psutil
 kornia>=0.7.1
 spandrel
 soundfile
-av>=14.1.0
+av>=14.2.0
 pydantic~=2.0