Merge dac0710c88 into 593be209a4

* feat: add essentials_category to nodes and blueprints for Essentials tab

Add ESSENTIALS_CATEGORY or essentials_category to 12 node classes and all
36 blueprint JSONs. Update SubgraphEntry TypedDict and subgraph_manager to
extract and pass through the field.

Fixes COM-15221

Amp-Thread-ID: https://ampcode.com/threads/T-019c83de-f7ab-7779-a451-0ba5940b56a9

* fix: import NotRequired from typing_extensions for Python 3.10 compat

* refactor: keep only node class ESSENTIALS_CATEGORY, remove blueprint/subgraph changes

Frontend will own blueprint categorization separately.

* fix: remove essentials_category from CreateVideo (not in spec)

---------

Co-authored-by: guill <jacob.e.segal@gmail.com>

.dockerignore

@@ -0,0 +1,8 @@
+.*
+!.git
+__pycache__/
+*.py[cod]
+input
+models
+notebooks
+output

.github/workflows/docker.yml

@@ -0,0 +1,101 @@
+name: Build and publish Docker images
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    branches:
+      - master
+  release:
+    types: [published]
+
+jobs:
+  docker:
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - id: cu118
+            name: CUDA 11.8
+            pytorch_install_args: "--index-url https://download.pytorch.org/whl/cu118"
+          - id: cu121
+            name: CUDA 12.1
+            pytorch_install_args: "--index-url https://download.pytorch.org/whl/cu121"
+          - id: cu124
+            name: CUDA 12.4
+            pytorch_install_args: "--index-url https://download.pytorch.org/whl/cu124"
+          - id: rocm6.2
+            name: ROCm 6.2
+            pytorch_install_args: "--index-url https://download.pytorch.org/whl/rocm6.2"
+          - id: cpu
+            name: CPU only
+            pytorch_install_args: "--index-url https://download.pytorch.org/whl/cpu"
+            extra_args: --cpu
+
+
+    name: ${{ matrix.name }}
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Check which repositories to use
+        id: repositories
+        run: |
+          echo "GHCR_IMAGE_NAME=ghcr.io/${GITHUB_REPOSITORY_OWNER}/comfyui" >> "$GITHUB_ENV"
+          if [[ -n "${DOCKERHUB_USERNAME}" ]]; then
+            echo "DOCKERHUB_IMAGE_NAME=${DOCKERHUB_USERNAME}/comfyui" >> "$GITHUB_ENV"
+          else
+            echo "DOCKERHUB_IMAGE_NAME=" >> "$GITHUB_ENV"
+            echo "No Docker Hub username set, only deploying to GitHub Container Repository"
+          fi
+        env:
+            DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_USERNAME }}
+            GITHUB_REPOSITORY_OWNER: ${{ github.repository_owner }}
+      - name: Docker meta
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          # list of Docker images to use as base name for tags
+          images: |
+            ${{ env.DOCKERHUB_IMAGE_NAME }}
+            ${{ env.GHCR_IMAGE_NAME }}
+          flavor: |
+            suffix=-${{ matrix.id }},onlatest=true
+          # generate Docker tags based on the following events/attributes
+          tags: |
+            type=ref,event=branch
+            type=ref,event=pr
+            type=semver,pattern={{version}}
+            type=semver,pattern={{major}}.{{minor}}
+            type=semver,pattern={{major}}
+            type=sha
+            # set latest tag for default branch
+            type=raw,value=latest,enable=${{ github.event_name == 'release' }}
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Login to Docker Hub
+        if: github.event_name != 'pull_request' && env.DOCKERHUB_IMAGE_NAME != ''
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+      - name: Login to GHCR
+        if: github.event_name != 'pull_request'
+        uses: docker/login-action@v3
+        with:
+          registry: ghcr.io
+          username: ${{ github.repository_owner }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+      - name: Build and push
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          push: ${{ github.event_name != 'pull_request' }}
+          tags: ${{ steps.meta.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels }}
+          build-args: |
+            PYTORCH_INSTALL_ARGS=${{ matrix.pytorch_install_args }}
+            EXTRA_ARGS=${{ matrix.extra_args }}
+          cache-from: type=gha,scope=${{ github.ref_name }}-${{ matrix.id }}
+          cache-to: type=gha,mode=max,scope=${{ github.ref_name }}-${{ matrix.id }}

Dockerfile

@@ -0,0 +1,79 @@
+# syntax=docker/dockerfile:1
+
+ARG PYTHON_VERSION=3.12
+
+FROM python:${PYTHON_VERSION}-slim
+
+ARG PYTORCH_INSTALL_ARGS=""
+ARG EXTRA_ARGS=""
+ARG USERNAME=comfyui
+ARG USER_UID=1000
+ARG USER_GID=${USER_UID}
+
+# Fail fast on errors or unset variables
+SHELL ["/bin/bash", "-eux", "-o", "pipefail", "-c"]
+
+RUN <<EOF
+	groupadd --gid ${USER_GID} ${USERNAME}
+	useradd --uid ${USER_UID} --gid ${USER_GID} -m ${USERNAME}
+EOF
+
+RUN <<EOF
+	apt-get update
+	apt-get install -y --no-install-recommends \
+		git \
+		git-lfs \
+		rsync \
+		fonts-recommended
+EOF
+
+# run instructions as user
+USER ${USER_UID}:${USER_GID}
+
+WORKDIR /app
+
+ENV XDG_CACHE_HOME=/cache
+ENV PIP_CACHE_DIR=/cache/pip
+ENV VIRTUAL_ENV=/app/venv
+ENV VIRTUAL_ENV_CUSTOM=/app/custom_venv
+
+# create cache directory. During build we will use a cache mount,
+# but later this is useful for custom node installs
+RUN --mount=type=cache,target=/cache/,uid=${USER_UID},gid=${USER_GID} \
+	mkdir -p ${PIP_CACHE_DIR}
+
+# create virtual environment to manage packages
+RUN python -m venv ${VIRTUAL_ENV}
+
+# run python from venv (prefer custom_venv over baked-in one)
+ENV PATH="${VIRTUAL_ENV_CUSTOM}/bin:${VIRTUAL_ENV}/bin:${PATH}"
+
+RUN --mount=type=cache,target=/cache/,uid=${USER_UID},gid=${USER_GID} \
+	pip install torch torchvision torchaudio ${PYTORCH_INSTALL_ARGS}
+
+# copy requirements files first so packages can be cached separately
+COPY --chown=${USER_UID}:${USER_GID} requirements.txt .
+RUN --mount=type=cache,target=/cache/,uid=${USER_UID},gid=${USER_GID} \
+	pip install -r requirements.txt
+
+# Not strictly required for comfyui, but prevents non-working variants of
+# cv2 being pulled in by custom nodes
+RUN --mount=type=cache,target=/cache/,uid=${USER_UID},gid=${USER_GID} \
+	pip install opencv-python-headless
+
+COPY --chown=${USER_UID}:${USER_GID} . .
+
+COPY --chown=nobody:${USER_GID} .git .git
+
+# default environment variables
+ENV COMFYUI_ADDRESS=0.0.0.0
+ENV COMFYUI_PORT=8188
+ENV COMFYUI_EXTRA_BUILD_ARGS="${EXTRA_ARGS}"
+ENV COMFYUI_EXTRA_ARGS=""
+# default start command
+CMD \
+	if [ -d "${VIRTUAL_ENV_CUSTOM}" ]; then \
+		rsync -aP "${VIRTUAL_ENV}/" "${VIRTUAL_ENV_CUSTOM}/" ;\
+		sed -i "s!${VIRTUAL_ENV}!${VIRTUAL_ENV_CUSTOM}!g" "${VIRTUAL_ENV_CUSTOM}/pyvenv.cfg" ;\
+	fi ;\
+	python -u main.py --listen "${COMFYUI_ADDRESS}" --port "${COMFYUI_PORT}" ${COMFYUI_EXTRA_BUILD_ARGS} ${COMFYUI_EXTRA_ARGS}

README.md

@@ -293,6 +293,45 @@ Install the dependencies by opening your terminal inside the ComfyUI folder and:
 
 After this you should have everything installed and can proceed to running ComfyUI.
 
+
+## Docker
+
+There are prebuilt docker images for AMD and NVIDIA GPUs on [GitHub Packages](https://ghcr.io/comfyanonymous/comfyui).
+
+You can pull them to your local docker registry with:
+
+```shell
+# For NVIDIA GPUs
+docker pull ghcr.io/comfyanonymous/comfyui:latest-cu124
+# For AMD GPUs
+docker pull ghcr.io/comfyanonymous/comfyui:latest-rocm6.2
+
+# For CPU only
+docker pull ghcr.io/comfyanonymous/comfyui:latest-cpu
+```
+
+### Building images manually
+
+You can build a docker image with the `Dockerfile` in this repo.
+Specify, `PYTORCH_INSTALL_ARGS` build arg with one of the PyTorch commands above to build for AMD or NVIDIA GPUs.
+```shell
+docker build --build-arg PYTORCH_INSTALL_ARGS="--index-url https://download.pytorch.org/whl/cu122" .
+```
+```shell
+docker build --build-arg PYTORCH_INSTALL_ARGS="--index-url https://download.pytorch.org/whl/rocm6.2" .
+```
+The `Dockerfile` requires BuildKit to be enabled. If your Docker does not support the buildx command, you can
+enable BuildKit by setting the `DOCKER_BUILDKIT` environment variable.
+```shell
+DOCKER_BUILDKIT=1 docker build --build-arg PYTORCH_INSTALL_ARGS="--index-url https://download.pytorch.org/whl/cu124" .
+```
+> [!NOTE]
+> For building the CPU-only image, it is recommended that you add the --cpu flag to the EXTRA_ARGS build arg:
+>
+> ```shell
+> docker build --build-arg PYTORCH_INSTALL_ARGS="--index-url https://download.pytorch.org/whl/cpu" --build-arg EXTRA_ARGS=--cpu .
+> ```
+
 ### Others:
 
 #### Apple Mac silicon

comfy/cli_args.py

@@ -83,6 +83,8 @@ fpte_group.add_argument("--fp16-text-enc", action="store_true", help="Store text
 fpte_group.add_argument("--fp32-text-enc", action="store_true", help="Store text encoder weights in fp32.")
 fpte_group.add_argument("--bf16-text-enc", action="store_true", help="Store text encoder weights in bf16.")
 
+parser.add_argument("--fp16-intermediates", action="store_true", help="Experimental: Use fp16 for intermediate tensors between nodes instead of fp32.")
+
 parser.add_argument("--force-channels-last", action="store_true", help="Force channels last format when inferencing the models.")
 
 parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE", const=-1, help="Use torch-directml.")

comfy/float.py

@@ -209,3 +209,39 @@ def stochastic_round_quantize_nvfp4_by_block(x, per_tensor_scale, pad_16x, seed=
         output_block[i:i + slice_size].copy_(block)
 
     return output_fp4, to_blocked(output_block, flatten=False)
+
+
+def stochastic_round_quantize_mxfp8_by_block(x, pad_32x, seed=0):
+    def roundup(x_val, multiple):
+        return ((x_val + multiple - 1) // multiple) * multiple
+
+    if pad_32x:
+        rows, cols = x.shape
+        padded_rows = roundup(rows, 32)
+        padded_cols = roundup(cols, 32)
+        if padded_rows != rows or padded_cols != cols:
+            x = torch.nn.functional.pad(x, (0, padded_cols - cols, 0, padded_rows - rows))
+
+    F8_E4M3_MAX = 448.0
+    E8M0_BIAS = 127
+    BLOCK_SIZE = 32
+
+    rows, cols = x.shape
+    x_blocked = x.reshape(rows, -1, BLOCK_SIZE)
+    max_abs = torch.amax(torch.abs(x_blocked), dim=-1)
+
+    # E8M0 block scales (power-of-2 exponents)
+    scale_needed = torch.clamp(max_abs.float() / F8_E4M3_MAX, min=2**(-127))
+    exp_biased = torch.clamp(torch.ceil(torch.log2(scale_needed)).to(torch.int32) + E8M0_BIAS, 0, 254)
+    block_scales_e8m0 = exp_biased.to(torch.uint8)
+
+    zero_mask = (max_abs == 0)
+    block_scales_f32 = (block_scales_e8m0.to(torch.int32) << 23).view(torch.float32)
+    block_scales_f32 = torch.where(zero_mask, torch.ones_like(block_scales_f32), block_scales_f32)
+
+    # Scale per-block then stochastic round
+    data_scaled = (x_blocked.float() / block_scales_f32.unsqueeze(-1)).reshape(rows, cols)
+    output_fp8 = stochastic_rounding(data_scaled, torch.float8_e4m3fn, seed=seed)
+
+    block_scales_e8m0 = torch.where(zero_mask, torch.zeros_like(block_scales_e8m0), block_scales_e8m0)
+    return output_fp8, to_blocked(block_scales_e8m0, flatten=False).view(torch.float8_e8m0fnu)

comfy/ldm/lightricks/vae/causal_video_autoencoder.py

@@ -11,6 +11,7 @@ from .causal_conv3d import CausalConv3d
 from .pixel_norm import PixelNorm
 from ..model import PixArtAlphaCombinedTimestepSizeEmbeddings
 import comfy.ops
+import comfy.model_management
 from comfy.ldm.modules.diffusionmodules.model import torch_cat_if_needed
 
 ops = comfy.ops.disable_weight_init
@@ -536,7 +537,7 @@ class Decoder(nn.Module):
                     mark_conv3d_ended(self.conv_out)
                 sample = self.conv_out(sample, causal=self.causal)
                 if sample is not None and sample.shape[2] > 0:
-                    output.append(sample)
+                    output.append(sample.to(comfy.model_management.intermediate_device()))
                 return
 
             up_block = self.up_blocks[idx]

comfy/model_management.py

@@ -400,7 +400,7 @@ try:
         if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
             if aotriton_supported(arch):  # AMD efficient attention implementation depends on aotriton.
                 if torch_version_numeric >= (2, 7):  # works on 2.6 but doesn't actually seem to improve much
-                    if any((a in arch) for a in ["gfx90a", "gfx942", "gfx950", "gfx1100", "gfx1101", "gfx1151"]):  # TODO: more arches, TODO: gfx950
+                    if any((a in arch) for a in ["gfx90a", "gfx942", "gfx950", "gfx1100", "gfx1101", "gfx1150", "gfx1151"]):  # TODO: more arches, TODO: gfx950
                         ENABLE_PYTORCH_ATTENTION = True
                 if rocm_version >= (7, 0):
                    if any((a in arch) for a in ["gfx1200", "gfx1201"]):
@@ -1050,6 +1050,12 @@ def intermediate_device():
     else:
         return torch.device("cpu")
 
+def intermediate_dtype():
+    if args.fp16_intermediates:
+        return torch.float16
+    else:
+        return torch.float32
+
 def vae_device():
     if args.cpu_vae:
         return torch.device("cpu")
@@ -1712,6 +1718,19 @@ def supports_nvfp4_compute(device=None):
 
     return True
 
+def supports_mxfp8_compute(device=None):
+    if not is_nvidia():
+        return False
+
+    if torch_version_numeric < (2, 10):
+        return False
+
+    props = torch.cuda.get_device_properties(device)
+    if props.major < 10:
+        return False
+
+    return True
+
 def extended_fp16_support():
     # TODO: check why some models work with fp16 on newer torch versions but not on older
     if torch_version_numeric < (2, 7):

comfy/ops.py

@@ -336,7 +336,10 @@ class disable_weight_init:
     class Linear(torch.nn.Linear, CastWeightBiasOp):
 
         def __init__(self, in_features, out_features, bias=True, device=None, dtype=None):
-            if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
+            # don't trust subclasses that BYO state dict loader to call us.
+            if (not comfy.model_management.WINDOWS
+                or not comfy.memory_management.aimdo_enabled
+                or type(self)._load_from_state_dict is not disable_weight_init.Linear._load_from_state_dict):
                 super().__init__(in_features, out_features, bias, device, dtype)
                 return
 
@@ -357,7 +360,9 @@ class disable_weight_init:
         def _load_from_state_dict(self, state_dict, prefix, local_metadata,
                                 strict, missing_keys, unexpected_keys, error_msgs):
 
-            if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
+            if (not comfy.model_management.WINDOWS
+                or not comfy.memory_management.aimdo_enabled
+                or type(self)._load_from_state_dict is not disable_weight_init.Linear._load_from_state_dict):
                 return super()._load_from_state_dict(state_dict, prefix, local_metadata, strict,
                                                      missing_keys, unexpected_keys, error_msgs)
             disable_weight_init._lazy_load_from_state_dict(
@@ -564,7 +569,10 @@ class disable_weight_init:
         def __init__(self, num_embeddings, embedding_dim, padding_idx=None, max_norm=None,
                      norm_type=2.0, scale_grad_by_freq=False, sparse=False, _weight=None,
                      _freeze=False, device=None, dtype=None):
-            if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
+            # don't trust subclasses that BYO state dict loader to call us.
+            if (not comfy.model_management.WINDOWS
+                or not comfy.memory_management.aimdo_enabled
+                or type(self)._load_from_state_dict is not disable_weight_init.Embedding._load_from_state_dict):
                 super().__init__(num_embeddings, embedding_dim, padding_idx, max_norm,
                                  norm_type, scale_grad_by_freq, sparse, _weight,
                                  _freeze, device, dtype)
@@ -590,7 +598,9 @@ class disable_weight_init:
         def _load_from_state_dict(self, state_dict, prefix, local_metadata,
                                 strict, missing_keys, unexpected_keys, error_msgs):
 
-            if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
+            if (not comfy.model_management.WINDOWS
+                or not comfy.memory_management.aimdo_enabled
+                or type(self)._load_from_state_dict is not disable_weight_init.Embedding._load_from_state_dict):
                 return super()._load_from_state_dict(state_dict, prefix, local_metadata, strict,
                                                      missing_keys, unexpected_keys, error_msgs)
             disable_weight_init._lazy_load_from_state_dict(
@@ -857,6 +867,22 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                             orig_shape=(self.out_features, self.in_features),
                         )
 
+                    elif self.quant_format == "mxfp8":
+                        # MXFP8: E8M0 block scales stored as uint8 in safetensors
+                        block_scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys,
+                                                             dtype=torch.uint8)
+
+                        if block_scale is None:
+                            raise ValueError(f"Missing MXFP8 block scales for layer {layer_name}")
+
+                        block_scale = block_scale.view(torch.float8_e8m0fnu)
+
+                        params = layout_cls.Params(
+                            scale=block_scale,
+                            orig_dtype=MixedPrecisionOps._compute_dtype,
+                            orig_shape=(self.out_features, self.in_features),
+                        )
+
                     elif self.quant_format == "nvfp4":
                         # NVFP4: tensor_scale (weight_scale_2) + block_scale (weight_scale)
                         tensor_scale = self._load_scale_param(state_dict, prefix, "weight_scale_2", device, manually_loaded_keys)
@@ -1006,12 +1032,15 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
 def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, model_config=None):
     fp8_compute = comfy.model_management.supports_fp8_compute(load_device) # TODO: if we support more ops this needs to be more granular
     nvfp4_compute = comfy.model_management.supports_nvfp4_compute(load_device)
+    mxfp8_compute = comfy.model_management.supports_mxfp8_compute(load_device)
 
     if model_config and hasattr(model_config, 'quant_config') and model_config.quant_config:
         logging.info("Using mixed precision operations")
         disabled = set()
         if not nvfp4_compute:
             disabled.add("nvfp4")
+        if not mxfp8_compute:
+            disabled.add("mxfp8")
         if not fp8_compute:
             disabled.add("float8_e4m3fn")
             disabled.add("float8_e5m2")

comfy/quant_ops.py

@@ -43,6 +43,18 @@ except ImportError as e:
     def get_layout_class(name):
         return None
 
+_CK_MXFP8_AVAILABLE = False
+if _CK_AVAILABLE:
+    try:
+        from comfy_kitchen.tensor import TensorCoreMXFP8Layout as _CKMxfp8Layout
+        _CK_MXFP8_AVAILABLE = True
+    except ImportError:
+        logging.warning("comfy_kitchen does not support MXFP8, please update comfy_kitchen.")
+
+if not _CK_MXFP8_AVAILABLE:
+    class _CKMxfp8Layout:
+        pass
+
 import comfy.float
 
 # ==============================================================================
@@ -84,6 +96,31 @@ class _TensorCoreFP8LayoutBase(_CKFp8Layout):
         return qdata, params
 
 
+class TensorCoreMXFP8Layout(_CKMxfp8Layout):
+    @classmethod
+    def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False):
+        if tensor.dim() != 2:
+            raise ValueError(f"MXFP8 requires 2D tensor, got {tensor.dim()}D")
+
+        orig_dtype = tensor.dtype
+        orig_shape = tuple(tensor.shape)
+
+        padded_shape = cls.get_padded_shape(orig_shape)
+        needs_padding = padded_shape != orig_shape
+
+        if stochastic_rounding > 0:
+            qdata, block_scale = comfy.float.stochastic_round_quantize_mxfp8_by_block(tensor, pad_32x=needs_padding, seed=stochastic_rounding)
+        else:
+            qdata, block_scale = ck.quantize_mxfp8(tensor, pad_32x=needs_padding)
+
+        params = cls.Params(
+            scale=block_scale,
+            orig_dtype=orig_dtype,
+            orig_shape=orig_shape,
+        )
+        return qdata, params
+
+
 class TensorCoreNVFP4Layout(_CKNvfp4Layout):
     @classmethod
     def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False):
@@ -137,6 +174,8 @@ register_layout_class("TensorCoreFP8Layout", TensorCoreFP8Layout)
 register_layout_class("TensorCoreFP8E4M3Layout", TensorCoreFP8E4M3Layout)
 register_layout_class("TensorCoreFP8E5M2Layout", TensorCoreFP8E5M2Layout)
 register_layout_class("TensorCoreNVFP4Layout", TensorCoreNVFP4Layout)
+if _CK_MXFP8_AVAILABLE:
+    register_layout_class("TensorCoreMXFP8Layout", TensorCoreMXFP8Layout)
 
 QUANT_ALGOS = {
     "float8_e4m3fn": {
@@ -157,6 +196,14 @@ QUANT_ALGOS = {
     },
 }
 
+if _CK_MXFP8_AVAILABLE:
+    QUANT_ALGOS["mxfp8"] = {
+        "storage_t": torch.float8_e4m3fn,
+        "parameters": {"weight_scale", "input_scale"},
+        "comfy_tensor_layout": "TensorCoreMXFP8Layout",
+        "group_size": 32,
+    }
+
 
 # ==============================================================================
 # Re-exports for backward compatibility

comfy/sd.py

@@ -871,13 +871,16 @@ class VAE:
                 pixels = torch.nn.functional.pad(pixels, (0, self.output_channels - pixels.shape[-1]), mode=mode, value=value)
         return pixels
 
+    def vae_output_dtype(self):
+        return model_management.intermediate_dtype()
+
     def decode_tiled_(self, samples, tile_x=64, tile_y=64, overlap = 16):
         steps = samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x, tile_y, overlap)
         steps += samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x // 2, tile_y * 2, overlap)
         steps += samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x * 2, tile_y // 2, overlap)
         pbar = comfy.utils.ProgressBar(steps)
 
-        decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
+        decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         output = self.process_output(
             (comfy.utils.tiled_scale(samples, decode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = self.upscale_ratio, output_device=self.output_device, pbar = pbar) +
             comfy.utils.tiled_scale(samples, decode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = self.upscale_ratio, output_device=self.output_device, pbar = pbar) +
@@ -887,16 +890,16 @@ class VAE:
 
     def decode_tiled_1d(self, samples, tile_x=256, overlap=32):
         if samples.ndim == 3:
-            decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
+            decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         else:
             og_shape = samples.shape
             samples = samples.reshape((og_shape[0], og_shape[1] * og_shape[2], -1))
-            decode_fn = lambda a: self.first_stage_model.decode(a.reshape((-1, og_shape[1], og_shape[2], a.shape[-1])).to(self.vae_dtype).to(self.device)).float()
+            decode_fn = lambda a: self.first_stage_model.decode(a.reshape((-1, og_shape[1], og_shape[2], a.shape[-1])).to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
 
         return self.process_output(comfy.utils.tiled_scale_multidim(samples, decode_fn, tile=(tile_x,), overlap=overlap, upscale_amount=self.upscale_ratio, out_channels=self.output_channels, output_device=self.output_device))
 
     def decode_tiled_3d(self, samples, tile_t=999, tile_x=32, tile_y=32, overlap=(1, 8, 8)):
-        decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
+        decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         return self.process_output(comfy.utils.tiled_scale_multidim(samples, decode_fn, tile=(tile_t, tile_x, tile_y), overlap=overlap, upscale_amount=self.upscale_ratio, out_channels=self.output_channels, index_formulas=self.upscale_index_formula, output_device=self.output_device))
 
     def encode_tiled_(self, pixel_samples, tile_x=512, tile_y=512, overlap = 64):
@@ -905,7 +908,7 @@ class VAE:
         steps += pixel_samples.shape[0] * comfy.utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x * 2, tile_y // 2, overlap)
         pbar = comfy.utils.ProgressBar(steps)
 
-        encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
+        encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         samples = comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x, tile_y, overlap, upscale_amount = (1/self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device, pbar=pbar)
         samples += comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = (1/self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device, pbar=pbar)
         samples += comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = (1/self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device, pbar=pbar)
@@ -914,7 +917,7 @@ class VAE:
 
     def encode_tiled_1d(self, samples, tile_x=256 * 2048, overlap=64 * 2048):
         if self.latent_dim == 1:
-            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
+            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
             out_channels = self.latent_channels
             upscale_amount = 1 / self.downscale_ratio
         else:
@@ -923,7 +926,7 @@ class VAE:
             tile_x = tile_x // extra_channel_size
             overlap = overlap // extra_channel_size
             upscale_amount = 1 / self.downscale_ratio
-            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).reshape(1, out_channels, -1).float()
+            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).reshape(1, out_channels, -1).to(dtype=self.vae_output_dtype())
 
         out = comfy.utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_x,), overlap=overlap, upscale_amount=upscale_amount, out_channels=out_channels, output_device=self.output_device)
         if self.latent_dim == 1:
@@ -932,7 +935,7 @@ class VAE:
             return out.reshape(samples.shape[0], self.latent_channels, extra_channel_size, -1)
 
     def encode_tiled_3d(self, samples, tile_t=9999, tile_x=512, tile_y=512, overlap=(1, 64, 64)):
-        encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
+        encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         return comfy.utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_t, tile_x, tile_y), overlap=overlap, upscale_amount=self.downscale_ratio, out_channels=self.latent_channels, downscale=True, index_formulas=self.downscale_index_formula, output_device=self.output_device)
 
     def decode(self, samples_in, vae_options={}):
@@ -950,9 +953,9 @@ class VAE:
 
             for x in range(0, samples_in.shape[0], batch_number):
                 samples = samples_in[x:x+batch_number].to(self.vae_dtype).to(self.device)
-                out = self.process_output(self.first_stage_model.decode(samples, **vae_options).to(self.output_device).float())
+                out = self.process_output(self.first_stage_model.decode(samples, **vae_options).to(self.output_device).to(dtype=self.vae_output_dtype()))
                 if pixel_samples is None:
-                    pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device)
+                    pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
                 pixel_samples[x:x+batch_number] = out
         except Exception as e:
             model_management.raise_non_oom(e)
@@ -1025,9 +1028,9 @@ class VAE:
             samples = None
             for x in range(0, pixel_samples.shape[0], batch_number):
                 pixels_in = self.process_input(pixel_samples[x:x + batch_number]).to(self.vae_dtype).to(self.device)
-                out = self.first_stage_model.encode(pixels_in).to(self.output_device).float()
+                out = self.first_stage_model.encode(pixels_in).to(self.output_device).to(dtype=self.vae_output_dtype())
                 if samples is None:
-                    samples = torch.empty((pixel_samples.shape[0],) + tuple(out.shape[1:]), device=self.output_device)
+                    samples = torch.empty((pixel_samples.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
                 samples[x:x + batch_number] = out
 
         except Exception as e:

comfy_api_nodes/nodes_kling.py

@@ -1459,6 +1459,7 @@ class OmniProEditVideoNode(IO.ComfyNode):
             node_id="KlingOmniProEditVideoNode",
             display_name="Kling 3.0 Omni Edit Video",
             category="api node/video/Kling",
+            essentials_category="Video Generation",
             description="Edit an existing video with the latest model from Kling.",
             inputs=[
                 IO.Combo.Input("model_name", options=["kling-v3-omni", "kling-video-o1"]),

comfy_api_nodes/nodes_recraft.py

@@ -833,6 +833,7 @@ class RecraftVectorizeImageNode(IO.ComfyNode):
             node_id="RecraftVectorizeImageNode",
             display_name="Recraft Vectorize Image",
             category="api node/image/Recraft",
+            essentials_category="Image Tools",
             description="Generates SVG synchronously from an input image.",
             inputs=[
                 IO.Image.Input("image"),

comfy_extras/nodes_audio.py

@@ -19,6 +19,7 @@ class EmptyLatentAudio(IO.ComfyNode):
             node_id="EmptyLatentAudio",
             display_name="Empty Latent Audio",
             category="latent/audio",
+            essentials_category="Audio",
             inputs=[
                 IO.Float.Input("seconds", default=47.6, min=1.0, max=1000.0, step=0.1),
                 IO.Int.Input(
@@ -185,6 +186,7 @@ class SaveAudioMP3(IO.ComfyNode):
             search_aliases=["export mp3"],
             display_name="Save Audio (MP3)",
             category="audio",
+            essentials_category="Audio",
             inputs=[
                 IO.Audio.Input("audio"),
                 IO.String.Input("filename_prefix", default="audio/ComfyUI"),

comfy_extras/nodes_image_compare.py

@@ -14,6 +14,7 @@ class ImageCompare(IO.ComfyNode):
             display_name="Image Compare",
             description="Compares two images side by side with a slider.",
             category="image",
+            essentials_category="Image Tools",
             is_experimental=True,
             is_output_node=True,
             inputs=[

comfy_extras/nodes_images.py

@@ -58,6 +58,7 @@ class ImageCropV2(IO.ComfyNode):
             search_aliases=["trim"],
             display_name="Image Crop",
             category="image/transform",
+            essentials_category="Image Tools",
             inputs=[
                 IO.Image.Input("image"),
                 IO.BoundingBox.Input("crop_region", component="ImageCrop"),

comfy_extras/nodes_post_processing.py

@@ -21,6 +21,7 @@ class Blend(io.ComfyNode):
             node_id="ImageBlend",
             display_name="Image Blend",
             category="image/postprocessing",
+            essentials_category="Image Tools",
             inputs=[
                 io.Image.Input("image1"),
                 io.Image.Input("image2"),

docker-compose.yaml

@@ -0,0 +1,21 @@
+services:
+  comfyui:
+    user: "1000:1000"
+    build: .
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [gpu]
+    ports:
+      - "8188:8188"
+    volumes:
+      - "./models:/app/models"
+      - "./input:/app/input"
+      - "./temp:/app/output/temp"
+      - "./output:/app/output"
+      - "./user:/app/user"
+      - "./custom_venv:/app/custom_venv"
+      - "./custom_nodes:/app/custom_nodes"

manager_requirements.txt

@@ -1 +1 @@
-comfyui_manager==4.1b4
+comfyui_manager==4.1b5

middleware/cache_middleware.py

@@ -32,7 +32,7 @@ async def cache_control(
     )
 
     if request.path.endswith(".js") or request.path.endswith(".css") or is_entry_point:
-        response.headers.setdefault("Cache-Control", "no-cache")
+        response.headers.setdefault("Cache-Control", "no-store")
         return response
 
     # Early return for non-image files - no cache headers needed

nodes.py

@@ -81,6 +81,7 @@ class CLIPTextEncode(ComfyNodeABC):
 
 
 class ConditioningCombine:
+    ESSENTIALS_CATEGORY = "Image Generation"
     @classmethod
     def INPUT_TYPES(s):
         return {"required": {"conditioning_1": ("CONDITIONING", ), "conditioning_2": ("CONDITIONING", )}}
@@ -1211,9 +1212,6 @@ class GLIGENTextBoxApply:
         return (c, )
 
 class EmptyLatentImage:
-    def __init__(self):
-        self.device = comfy.model_management.intermediate_device()
-
     @classmethod
     def INPUT_TYPES(s):
         return {
@@ -1232,7 +1230,7 @@ class EmptyLatentImage:
     SEARCH_ALIASES = ["empty", "empty latent", "new latent", "create latent", "blank latent", "blank"]
 
     def generate(self, width, height, batch_size=1):
-        latent = torch.zeros([batch_size, 4, height // 8, width // 8], device=self.device)
+        latent = torch.zeros([batch_size, 4, height // 8, width // 8], device=comfy.model_management.intermediate_device(), dtype=comfy.model_management.intermediate_dtype())
         return ({"samples": latent, "downscale_ratio_spacial": 8}, )
 
 
@@ -1724,6 +1722,8 @@ class LoadImage:
         output_masks = []
         w, h = None, None
 
+        dtype = comfy.model_management.intermediate_dtype()
+
         for i in ImageSequence.Iterator(img):
             i = node_helpers.pillow(ImageOps.exif_transpose, i)
 
@@ -1748,8 +1748,8 @@ class LoadImage:
                 mask = 1. - torch.from_numpy(mask)
             else:
                 mask = torch.zeros((64,64), dtype=torch.float32, device="cpu")
-            output_images.append(image)
-            output_masks.append(mask.unsqueeze(0))
+            output_images.append(image.to(dtype=dtype))
+            output_masks.append(mask.unsqueeze(0).to(dtype=dtype))
 
             if img.format == "MPO":
                 break  # ignore all frames except the first one for MPO format
@@ -1779,6 +1779,7 @@ class LoadImage:
         return True
 
 class LoadImageMask:
+    ESSENTIALS_CATEGORY = "Image Tools"
     SEARCH_ALIASES = ["import mask", "alpha mask", "channel mask"]
 
     _color_channels = ["alpha", "red", "green", "blue"]
@@ -1887,6 +1888,7 @@ class ImageScale:
         return (s,)
 
 class ImageScaleBy:
+    ESSENTIALS_CATEGORY = "Image Tools"
     upscale_methods = ["nearest-exact", "bilinear", "area", "bicubic", "lanczos"]
 
     @classmethod

requirements.txt

@@ -1,4 +1,4 @@
-comfyui-frontend-package==1.41.19
+comfyui-frontend-package==1.41.20
 comfyui-workflow-templates==0.9.21
 comfyui-embedded-docs==0.4.3
 torch
@@ -23,7 +23,7 @@ SQLAlchemy
 filelock
 av>=14.2.0
 comfy-kitchen>=0.2.8
-comfy-aimdo>=0.2.11
+comfy-aimdo>=0.2.12
 requests
 simpleeval>=1.0.0
 blake3

server.py

@@ -310,7 +310,7 @@ class PromptServer():
         @routes.get("/")
         async def get_root(request):
             response = web.FileResponse(os.path.join(self.web_root, "index.html"))
-            response.headers['Cache-Control'] = 'no-cache'
+            response.headers['Cache-Control'] = 'no-store, must-revalidate'
             response.headers["Pragma"] = "no-cache"
             response.headers["Expires"] = "0"
             return response

tests-unit/server_test/test_cache_control.py

@@ -28,31 +28,31 @@ CACHE_SCENARIOS = [
     },
     # JavaScript/CSS scenarios
     {
-        "name": "js_no_cache",
+        "name": "js_no_store",
         "path": "/script.js",
         "status": 200,
-        "expected_cache": "no-cache",
+        "expected_cache": "no-store",
         "should_have_header": True,
     },
     {
-        "name": "css_no_cache",
+        "name": "css_no_store",
         "path": "/styles.css",
         "status": 200,
-        "expected_cache": "no-cache",
+        "expected_cache": "no-store",
         "should_have_header": True,
     },
     {
-        "name": "index_json_no_cache",
+        "name": "index_json_no_store",
         "path": "/api/index.json",
         "status": 200,
-        "expected_cache": "no-cache",
+        "expected_cache": "no-store",
         "should_have_header": True,
     },
     {
-        "name": "localized_index_json_no_cache",
+        "name": "localized_index_json_no_store",
         "path": "/templates/index.zh.json",
         "status": 200,
-        "expected_cache": "no-cache",
+        "expected_cache": "no-store",
         "should_have_header": True,
     },
     # Non-matching files