Merge b35b397720 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 +1,31 @@
+# This file should remain in sync with .gitignore. If you need to make changes,
+# please add a comment explaining why. For items that must be removed, comment
+# them out instead of deleting them.
+__pycache__/
+*.py[cod]
+/output/
+/input/
+# This file prevents the image from building and would be overwritten by the
+# /data volume in any case.
+#!/input/example.png
+/models/
+/temp/
+/custom_nodes/
+!custom_nodes/example_node.py.example
+extra_model_paths.yaml
+/.vs
+.vscode/
+.idea/
+venv/
+.venv/
+/web/extensions/*
+!/web/extensions/logging.js.example
+!/web/extensions/core/
+/tests-ui/data/object_info.json
+/user/
+*.log
+web_custom_versions/
+.DS_Store
+openapi.yaml
+filtered-openapi.yaml
+uv.lock

.gitattributes

@@ -1,3 +1,6 @@
 /web/assets/** linguist-generated
 /web/** linguist-vendored
 comfy_api_nodes/apis/__init__.py linguist-generated
+# Force LF eol for Docker entrypoint (fix "exec: no such file or directory"
+# error with CRLF checkouts)
+entrypoint.sh text eol=lf

.gitignore

@@ -1,3 +1,4 @@
+# If you modify this file, remember to update .dockerignore as well.
 __pycache__/
 *.py[cod]
 /output/

Dockerfile

@@ -0,0 +1,95 @@
+# Docker buildfile for the ComfyUI image, with support for hardware
+# acceleration, file ownership synchronization, custom nodes, and custom node
+# managers.
+
+# While Python 3.13 is well supported by ComfyUI, some older custom node packs
+# may not work correctly with this version, which is why we're staying on Python
+# 3.12 for now.
+#
+# Users are free to try different base Python image tags (e.g., 3.13, alpine,
+# *-slim), but for maintainability, only one base version is officially
+# supported at a time.
+FROM python:3.12.12-trixie
+
+# Install cmake, which is an indirect installation dependencies
+RUN apt-get update && apt-get install -y --no-install-recommends cmake
+
+# Create a regular user whose UID and GID will match the host user's at runtime.
+# Also create a home directory for this user (-m), as some common Python tools
+# (such as uv) interact with the user’s home directory.
+RUN useradd -m comfyui
+
+# Install ComfyUI under /comfyui and set folder ownership to the comfyui user.
+# With the legacy Docker builder (DOCKER_BUILDKIT=0), WORKDIR always creates missing
+# directories as root (even if a different USER is active). To ensure the comfyui user
+# can write inside, ownership must be fixed manually.
+WORKDIR /comfyui
+RUN chown comfyui:comfyui .
+
+# Install ComfyUI as ComfyUI
+USER comfyui
+
+# Set up a Python virtual environment and configure it as the default Python.
+#
+# Reasons for using a virtual environment:
+# - Some custom nodes use third-party tools like uv, which do not support
+#   user-level installations.
+# - Custom node managers may install or update dependencies as the regular user,
+#   so a global installation is not an option.
+# This leaves virtual environments as the only viable choice.
+RUN python -m venv .venv
+ENV PATH="/comfyui/.venv/bin:$PATH"
+
+# Take PIP_EXTRA_INDEX_URL as a required build argument and bake it into the
+# image as an environment variable. This variable is then used to select
+# hardware-specific flavors of PyTorch, enabling support for CPU, AMD, and newer
+# CUDA inference.
+ARG PIP_EXTRA_INDEX_URL
+ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
+
+# Install ComfyUI's Python dependencies. Although dependency keeping is also
+# performed at startup, building ComfyUI's base dependencies into the image
+# significantly speeds up each containers' first run.
+#
+# Since this step takes a long time to complete, it's performed early to take
+# advantage of Docker's build cache, thereby accelerating subsequent builds.
+COPY requirements.txt manager_requirements.txt ./
+RUN pip install                                \
+    --no-cache-dir                             \
+    --disable-pip-version-check                \
+    --extra-index-url "${PIP_EXTRA_INDEX_URL}" \
+    -r requirements.txt
+
+# Install ComfyUI
+COPY . .
+
+# Purely declarative: inform Docker and image users that this image is designed
+# to listen on port 8188 for the web GUI.
+EXPOSE 8188
+
+# Declare persistent volumes. We assign one volume per data directory to match
+# ComfyUI’s natural file layout and to let users selectively choose which
+# directories they want to mount.
+VOLUME /comfyui/.venv
+VOLUME /comfyui/custom_nodes
+VOLUME /comfyui/input
+VOLUME /comfyui/models
+VOLUME /comfyui/output
+VOLUME /comfyui/temp
+VOLUME /comfyui/user
+VOLUME /home/comfyui
+
+# Switch back to root to run the entrypoint and to install additional system
+# dependencies
+USER root
+
+# Configure entrypoint
+RUN chmod +x entrypoint.sh
+ENTRYPOINT [ "./entrypoint.sh" ]
+CMD [ "python", "./main.py" ]
+
+# Install additional system dependencies
+ARG APT_EXTRA_PACKAGES
+RUN apt-get install -y --no-install-recommends $APT_EXTRA_PACKAGES \
+	&& apt-get clean                                               \
+	&& rm -rf /var/lib/apt/lists/*

README.md

@@ -38,6 +38,8 @@ ComfyUI lets you design and execute advanced stable diffusion pipelines using a
 
 ## Get Started
 
+### Local
+
 #### [Desktop Application](https://www.comfy.org/download)
 - The easiest way to get started.
 - Available on Windows & macOS.
@@ -46,11 +48,22 @@ ComfyUI lets you design and execute advanced stable diffusion pipelines using a
 - Get the latest commits and completely portable.
 - Available on Windows.
 
+#### [Docker Install](#running-with-docker)
+- Run ComfyUI inside an isolated Docker container
+- Most secure way to run ComfyUI and custom node packs
+- Requires Docker and Docker Compose
+- Supports NVIDIA GPUs (Not tested on other hardware.)
+
 #### [Manual Install](#manual-install-windows-linux)
 Supports all operating systems and GPU types (NVIDIA, AMD, Intel, Apple Silicon, Ascend).
 
-## [Examples](https://comfyanonymous.github.io/ComfyUI_examples/)
-See what ComfyUI can do with the [example workflows](https://comfyanonymous.github.io/ComfyUI_examples/).
+### Cloud
+
+#### [Comfy Cloud](https://www.comfy.org/cloud)
+- Our official paid cloud version for those who can't afford local hardware.
+
+## Examples
+See what ComfyUI can do with the [newer template workflows](https://comfy.org/workflows) or old [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.
@@ -348,6 +361,31 @@ For models compatible with Iluvatar Extension for PyTorch. Here's a step-by-step
 | `--enable-manager-legacy-ui` | Use the legacy manager UI instead of the new UI (requires `--enable-manager`) |
 | `--disable-manager-ui` | Disable the manager UI and endpoints while keeping background features like security checks and scheduled installation completion (requires `--enable-manager`) |
 
+## Running with Docker
+
+Start by installing Docker, Docker Compose, and the NVIDIA Container Toolkit on
+your host. Next, edit `compose.yaml` and update the `UID` and `GID` variables to
+match your host user's UID and GID. Also ensure that `PIP_EXTRA_INDEX_URL`
+points to the recommended non-experimental index URL for your hardware, as
+described in the beginning of the [manual installation
+section](#manual-install-windows-linux). Additional fields are documented in the
+file for further customization.
+
+Once ready, build and run the image locally:
+
+```shell
+# (Re)build the Docker image. Run this before the first start, after updating
+# ComfyUI, or after changing any build arguments in `compose.yaml`.
+docker compose build
+# Start ComfyUI. This reuses the most recently built image.
+docker compose up
+```
+
+To stop and remove the container along with its volumes, run:
+
+```shell
+docker compose down -v
+```
 
 # Running
 

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/memory_management.py

@@ -1,9 +1,68 @@
 import math
+import ctypes
+import threading
+import dataclasses
 import torch
 from typing import NamedTuple
 
 from comfy.quant_ops import QuantizedTensor
 
+
+class TensorFileSlice(NamedTuple):
+    file_ref: object
+    thread_id: int
+    offset: int
+    size: int
+
+
+def read_tensor_file_slice_into(tensor, destination):
+
+    if isinstance(tensor, QuantizedTensor):
+        if not isinstance(destination, QuantizedTensor):
+            return False
+        if tensor._layout_cls != destination._layout_cls:
+            return False
+
+        if not read_tensor_file_slice_into(tensor._qdata, destination._qdata):
+            return False
+
+        dst_orig_dtype = destination._params.orig_dtype
+        destination._params.copy_from(tensor._params, non_blocking=False)
+        destination._params = dataclasses.replace(destination._params, orig_dtype=dst_orig_dtype)
+        return True
+
+    info = getattr(tensor.untyped_storage(), "_comfy_tensor_file_slice", None)
+    if info is None:
+        return False
+
+    file_obj = info.file_ref
+    if (destination.device.type != "cpu"
+            or file_obj is None
+            or threading.get_ident() != info.thread_id
+            or destination.numel() * destination.element_size() < info.size):
+        return False
+
+    if info.size == 0:
+        return True
+
+    buf_type = ctypes.c_ubyte * info.size
+    view = memoryview(buf_type.from_address(destination.data_ptr()))
+
+    try:
+        file_obj.seek(info.offset)
+        done = 0
+        while done < info.size:
+            try:
+                n = file_obj.readinto(view[done:])
+            except OSError:
+                return False
+            if n <= 0:
+                return False
+            done += n
+        return True
+    finally:
+        view.release()
+
 class TensorGeometry(NamedTuple):
     shape: any
     dtype: torch.dtype

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"]):
@@ -505,6 +505,28 @@ def module_size(module):
         module_mem += t.nbytes
     return module_mem
 
+def module_mmap_residency(module, free=False):
+    mmap_touched_mem = 0
+    module_mem = 0
+    bounced_mmaps = set()
+    sd = module.state_dict()
+    for k in sd:
+        t = sd[k]
+        module_mem += t.nbytes
+        storage = t._qdata.untyped_storage() if isinstance(t, comfy.quant_ops.QuantizedTensor) else t.untyped_storage()
+        if not getattr(storage, "_comfy_tensor_mmap_touched", False):
+            continue
+        mmap_touched_mem += t.nbytes
+        if not free:
+            continue
+        storage._comfy_tensor_mmap_touched = False
+        mmap_obj = storage._comfy_tensor_mmap_refs[0]
+        if mmap_obj in bounced_mmaps:
+            continue
+        mmap_obj.bounce()
+        bounced_mmaps.add(mmap_obj)
+    return mmap_touched_mem, module_mem
+
 class LoadedModel:
     def __init__(self, model):
         self._set_model(model)
@@ -532,6 +554,9 @@ class LoadedModel:
     def model_memory(self):
         return self.model.model_size()
 
+    def model_mmap_residency(self, free=False):
+        return self.model.model_mmap_residency(free=free)
+
     def model_loaded_memory(self):
         return self.model.loaded_size()
 
@@ -633,7 +658,7 @@ def extra_reserved_memory():
 def minimum_inference_memory():
     return (1024 * 1024 * 1024) * 0.8 + extra_reserved_memory()
 
-def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_required=0):
+def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, pins_required=0, ram_required=0):
     cleanup_models_gc()
     unloaded_model = []
     can_unload = []
@@ -646,13 +671,14 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_
                 can_unload.append((-shift_model.model_offloaded_memory(), sys.getrefcount(shift_model.model), shift_model.model_memory(), i))
                 shift_model.currently_used = False
 
-    for x in sorted(can_unload):
+    can_unload_sorted = sorted(can_unload)
+    for x in can_unload_sorted:
         i = x[-1]
         memory_to_free = 1e32
-        ram_to_free = 1e32
+        pins_to_free = 1e32
         if not DISABLE_SMART_MEMORY:
             memory_to_free = memory_required - get_free_memory(device)
-            ram_to_free = ram_required - get_free_ram()
+            pins_to_free = pins_required - get_free_ram()
             if current_loaded_models[i].model.is_dynamic() and for_dynamic:
                 #don't actually unload dynamic models for the sake of other dynamic models
                 #as that works on-demand.
@@ -661,9 +687,18 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_
         if memory_to_free > 0 and current_loaded_models[i].model_unload(memory_to_free):
             logging.debug(f"Unloading {current_loaded_models[i].model.model.__class__.__name__}")
             unloaded_model.append(i)
-        if ram_to_free > 0:
+        if pins_to_free > 0:
+            logging.debug(f"PIN Unloading {current_loaded_models[i].model.model.__class__.__name__}")
+            current_loaded_models[i].model.partially_unload_ram(pins_to_free)
+
+    for x in can_unload_sorted:
+        i = x[-1]
+        ram_to_free = ram_required - psutil.virtual_memory().available
+        if ram_to_free <= 0 and i not in unloaded_model:
+            continue
+        resident_memory, _ = current_loaded_models[i].model_mmap_residency(free=True)
+        if resident_memory > 0:
             logging.debug(f"RAM Unloading {current_loaded_models[i].model.model.__class__.__name__}")
-            current_loaded_models[i].model.partially_unload_ram(ram_to_free)
 
     for i in sorted(unloaded_model, reverse=True):
         unloaded_models.append(current_loaded_models.pop(i))
@@ -729,17 +764,27 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
 
 
     total_memory_required = {}
+    total_pins_required = {}
     total_ram_required = {}
     for loaded_model in models_to_load:
-        total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device)
-        #x2, one to make sure the OS can fit the model for loading in disk cache, and for us to do any pinning we
-        #want to do.
-        #FIXME: This should subtract off the to_load current pin consumption.
-        total_ram_required[loaded_model.device] = total_ram_required.get(loaded_model.device, 0) + loaded_model.model_memory() * 2
+        device = loaded_model.device
+        total_memory_required[device] = total_memory_required.get(device, 0) + loaded_model.model_memory_required(device)
+        resident_memory, model_memory = loaded_model.model_mmap_residency()
+        pinned_memory = loaded_model.model.pinned_memory_size()
+        #FIXME: This can over-free the pins as it budgets to pin the entire model. We should
+        #make this JIT to keep as much pinned as possible.
+        pins_required = model_memory - pinned_memory
+        ram_required = model_memory - resident_memory
+        total_pins_required[device] = total_pins_required.get(device, 0) + pins_required
+        total_ram_required[device] = total_ram_required.get(device, 0) + ram_required
 
     for device in total_memory_required:
         if device != torch.device("cpu"):
-            free_memory(total_memory_required[device] * 1.1 + extra_mem, device, for_dynamic=free_for_dynamic, ram_required=total_ram_required[device])
+            free_memory(total_memory_required[device] * 1.1 + extra_mem,
+                        device,
+                        for_dynamic=free_for_dynamic,
+                        pins_required=total_pins_required[device],
+                        ram_required=total_ram_required[device])
 
     for device in total_memory_required:
         if device != torch.device("cpu"):
@@ -1005,6 +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")
@@ -1225,6 +1276,11 @@ def cast_to_gathered(tensors, r, non_blocking=False, stream=None):
             dest_view = dest_views.pop(0)
             if tensor is None:
                 continue
+            if comfy.memory_management.read_tensor_file_slice_into(tensor, dest_view):
+                continue
+            storage = tensor._qdata.untyped_storage() if isinstance(tensor, comfy.quant_ops.QuantizedTensor) else tensor.untyped_storage()
+            if hasattr(storage, "_comfy_tensor_mmap_touched"):
+                storage._comfy_tensor_mmap_touched = True
             dest_view.copy_(tensor, non_blocking=non_blocking)
 
 
@@ -1662,6 +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/model_patcher.py

@@ -297,6 +297,9 @@ class ModelPatcher:
         self.size = comfy.model_management.module_size(self.model)
         return self.size
 
+    def model_mmap_residency(self, free=False):
+        return comfy.model_management.module_mmap_residency(self.model, free=free)
+
     def get_ram_usage(self):
         return self.model_size()
 
@@ -1063,6 +1066,10 @@ class ModelPatcher:
 
             return self.model.model_loaded_weight_memory - current_used
 
+    def pinned_memory_size(self):
+        # Pinned memory pressure tracking is only implemented for DynamicVram loading
+        return 0
+
     def partially_unload_ram(self, ram_to_unload):
         pass
 
@@ -1653,6 +1660,16 @@ class ModelPatcherDynamic(ModelPatcher):
 
         return freed
 
+    def pinned_memory_size(self):
+        total = 0
+        loading = self._load_list(for_dynamic=True)
+        for x in loading:
+            _, _, _, _, m, _ = x
+            pin = comfy.pinned_memory.get_pin(m)
+            if pin is not None:
+                total += pin.numel() * pin.element_size()
+        return total
+
     def partially_unload_ram(self, ram_to_unload):
         loading = self._load_list(for_dynamic=True, default_device=self.offload_device)
         for x in loading:

comfy/ops.py

@@ -306,10 +306,40 @@ class CastWeightBiasOp:
     bias_function = []
 
 class disable_weight_init:
+    @staticmethod
+    def _lazy_load_from_state_dict(module, state_dict, prefix, local_metadata,
+                                   missing_keys, unexpected_keys, weight_shape,
+                                   bias_shape=None):
+        assign_to_params_buffers = local_metadata.get("assign_to_params_buffers", False)
+        prefix_len = len(prefix)
+        for k, v in state_dict.items():
+            key = k[prefix_len:]
+            if key == "weight":
+                if not assign_to_params_buffers:
+                    v = v.clone()
+                module.weight = torch.nn.Parameter(v, requires_grad=False)
+            elif bias_shape is not None and key == "bias" and v is not None:
+                if not assign_to_params_buffers:
+                    v = v.clone()
+                module.bias = torch.nn.Parameter(v, requires_grad=False)
+            else:
+                unexpected_keys.append(k)
+
+        if module.weight is None:
+            module.weight = torch.nn.Parameter(torch.zeros(weight_shape), requires_grad=False)
+            missing_keys.append(prefix + "weight")
+
+        if bias_shape is not None and module.bias is None and getattr(module, "comfy_need_lazy_init_bias", False):
+            module.bias = torch.nn.Parameter(torch.zeros(bias_shape), requires_grad=False)
+            missing_keys.append(prefix + "bias")
+
     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
 
@@ -330,32 +360,21 @@ 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)
-            assign_to_params_buffers = local_metadata.get("assign_to_params_buffers", False)
-            prefix_len = len(prefix)
-            for k,v in state_dict.items():
-                if k[prefix_len:] == "weight":
-                    if not assign_to_params_buffers:
-                        v = v.clone()
-                    self.weight = torch.nn.Parameter(v, requires_grad=False)
-                elif k[prefix_len:] == "bias" and v is not None:
-                    if not assign_to_params_buffers:
-                        v = v.clone()
-                    self.bias = torch.nn.Parameter(v, requires_grad=False)
-                else:
-                    unexpected_keys.append(k)
-
-            #Reconcile default construction of the weight if its missing.
-            if self.weight is None:
-                v = torch.zeros(self.in_features, self.out_features)
-                self.weight = torch.nn.Parameter(v, requires_grad=False)
-                missing_keys.append(prefix+"weight")
-            if self.bias is None and self.comfy_need_lazy_init_bias:
-                v = torch.zeros(self.out_features,)
-                self.bias = torch.nn.Parameter(v, requires_grad=False)
-                missing_keys.append(prefix+"bias")
+            disable_weight_init._lazy_load_from_state_dict(
+                self,
+                state_dict,
+                prefix,
+                local_metadata,
+                missing_keys,
+                unexpected_keys,
+                weight_shape=(self.in_features, self.out_features),
+                bias_shape=(self.out_features,),
+            )
 
 
         def reset_parameters(self):
@@ -547,6 +566,53 @@ class disable_weight_init:
                 return super().forward(*args, **kwargs)
 
     class Embedding(torch.nn.Embedding, CastWeightBiasOp):
+        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):
+            # 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)
+                return
+
+            torch.nn.Module.__init__(self)
+            self.num_embeddings = num_embeddings
+            self.embedding_dim = embedding_dim
+            self.padding_idx = padding_idx
+            self.max_norm = max_norm
+            self.norm_type = norm_type
+            self.scale_grad_by_freq = scale_grad_by_freq
+            self.sparse = sparse
+            # Keep shape/dtype visible for module introspection without reserving storage.
+            embedding_dtype = dtype if dtype is not None else torch.get_default_dtype()
+            self.weight = torch.nn.Parameter(
+                torch.empty((num_embeddings, embedding_dim), device="meta", dtype=embedding_dtype),
+                requires_grad=False,
+            )
+            self.bias = None
+            self.weight_comfy_model_dtype = dtype
+
+        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
+                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(
+                self,
+                state_dict,
+                prefix,
+                local_metadata,
+                missing_keys,
+                unexpected_keys,
+                weight_shape=(self.num_embeddings, self.embedding_dim),
+            )
+
         def reset_parameters(self):
             self.bias = None
             return None
@@ -801,6 +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)
@@ -950,12 +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/pinned_memory.py

@@ -1,6 +1,7 @@
-import torch
 import comfy.model_management
 import comfy.memory_management
+import comfy_aimdo.host_buffer
+import comfy_aimdo.torch
 
 from comfy.cli_args import args
 
@@ -12,18 +13,31 @@ def pin_memory(module):
         return
     #FIXME: This is a RAM cache trigger event
     size = comfy.memory_management.vram_aligned_size([ module.weight, module.bias ])
-    pin = torch.empty((size,), dtype=torch.uint8)
-    if comfy.model_management.pin_memory(pin):
-        module._pin = pin
-    else:
+
+    if comfy.model_management.MAX_PINNED_MEMORY <= 0 or (comfy.model_management.TOTAL_PINNED_MEMORY + size) > comfy.model_management.MAX_PINNED_MEMORY:
         module.pin_failed = True
         return False
+
+    try:
+        hostbuf = comfy_aimdo.host_buffer.HostBuffer(size)
+    except RuntimeError:
+        module.pin_failed = True
+        return False
+
+    module._pin = comfy_aimdo.torch.hostbuf_to_tensor(hostbuf)
+    module._pin_hostbuf = hostbuf
+    comfy.model_management.TOTAL_PINNED_MEMORY += size
     return True
 
 def unpin_memory(module):
     if get_pin(module) is None:
         return 0
     size = module._pin.numel() * module._pin.element_size()
-    comfy.model_management.unpin_memory(module._pin)
+
+    comfy.model_management.TOTAL_PINNED_MEMORY -= size
+    if comfy.model_management.TOTAL_PINNED_MEMORY < 0:
+        comfy.model_management.TOTAL_PINNED_MEMORY = 0
+
     del module._pin
+    del module._pin_hostbuf
     return size

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/utils.py

@@ -20,6 +20,8 @@
 import torch
 import math
 import struct
+import ctypes
+import os
 import comfy.memory_management
 import safetensors.torch
 import numpy as np
@@ -32,7 +34,7 @@ from einops import rearrange
 from comfy.cli_args import args
 import json
 import time
-import mmap
+import threading
 import warnings
 
 MMAP_TORCH_FILES = args.mmap_torch_files
@@ -81,14 +83,17 @@ _TYPES = {
 }
 
 def load_safetensors(ckpt):
-    f = open(ckpt, "rb")
-    mapping = mmap.mmap(f.fileno(), 0, access=mmap.ACCESS_READ)
-    mv = memoryview(mapping)
+    import comfy_aimdo.model_mmap
 
-    header_size = struct.unpack("<Q", mapping[:8])[0]
-    header = json.loads(mapping[8:8+header_size].decode("utf-8"))
+    f = open(ckpt, "rb", buffering=0)
+    model_mmap = comfy_aimdo.model_mmap.ModelMMAP(ckpt)
+    file_size = os.path.getsize(ckpt)
+    mv = memoryview((ctypes.c_uint8 * file_size).from_address(model_mmap.get()))
 
-    mv = mv[8 + header_size:]
+    header_size = struct.unpack("<Q", mv[:8])[0]
+    header = json.loads(mv[8:8 + header_size].tobytes().decode("utf-8"))
+
+    mv = mv[(data_base_offset := 8 + header_size):]
 
     sd = {}
     for name, info in header.items():
@@ -102,7 +107,14 @@ def load_safetensors(ckpt):
             with warnings.catch_warnings():
                 #We are working with read-only RAM by design
                 warnings.filterwarnings("ignore", message="The given buffer is not writable")
-                sd[name] = torch.frombuffer(mv[start:end], dtype=_TYPES[info["dtype"]]).view(info["shape"])
+                tensor = torch.frombuffer(mv[start:end], dtype=_TYPES[info["dtype"]]).view(info["shape"])
+                storage = tensor.untyped_storage()
+                setattr(storage,
+                        "_comfy_tensor_file_slice",
+                        comfy.memory_management.TensorFileSlice(f, threading.get_ident(), data_base_offset + start, end - start))
+                setattr(storage, "_comfy_tensor_mmap_refs", (model_mmap, mv))
+                setattr(storage, "_comfy_tensor_mmap_touched", False)
+                sd[name] = tensor
 
     return sd, header.get("__metadata__", {}),
 

comfy_api_nodes/nodes_hunyuan3d.py

@@ -1,3 +1,7 @@
+import zipfile
+from io import BytesIO
+
+import torch
 from typing_extensions import override
 
 from comfy_api.latest import IO, ComfyExtension, Input, Types
@@ -17,7 +21,10 @@ from comfy_api_nodes.apis.hunyuan3d import (
 )
 from comfy_api_nodes.util import (
     ApiEndpoint,
+    bytesio_to_image_tensor,
+    download_url_to_bytesio,
     download_url_to_file_3d,
+    download_url_to_image_tensor,
     downscale_image_tensor_by_max_side,
     poll_op,
     sync_op,
@@ -36,6 +43,68 @@ def _is_tencent_rate_limited(status: int, body: object) -> bool:
     )
 
 
+class ObjZipResult:
+    __slots__ = ("obj", "texture", "metallic", "normal", "roughness")
+
+    def __init__(
+        self,
+        obj: Types.File3D,
+        texture: Input.Image | None = None,
+        metallic: Input.Image | None = None,
+        normal: Input.Image | None = None,
+        roughness: Input.Image | None = None,
+    ):
+        self.obj = obj
+        self.texture = texture
+        self.metallic = metallic
+        self.normal = normal
+        self.roughness = roughness
+
+
+async def download_and_extract_obj_zip(url: str) -> ObjZipResult:
+    """The Tencent API returns OBJ results as ZIP archives containing the .obj mesh, and texture images.
+
+    When PBR is enabled, the ZIP may contain additional metallic, normal, and roughness maps
+    identified by their filename suffixes.
+    """
+    data = BytesIO()
+    await download_url_to_bytesio(url, data)
+    data.seek(0)
+    if not zipfile.is_zipfile(data):
+        data.seek(0)
+        return ObjZipResult(obj=Types.File3D(source=data, file_format="obj"))
+    data.seek(0)
+    obj_bytes = None
+    textures: dict[str, Input.Image] = {}
+    with zipfile.ZipFile(data) as zf:
+        for name in zf.namelist():
+            lower = name.lower()
+            if lower.endswith(".obj"):
+                obj_bytes = zf.read(name)
+            elif any(lower.endswith(ext) for ext in (".png", ".jpg", ".jpeg", ".bmp", ".tiff", ".webp")):
+                stem = lower.rsplit(".", 1)[0]
+                tensor = bytesio_to_image_tensor(BytesIO(zf.read(name)), mode="RGB")
+                matched_key = "texture"
+                for suffix, key in {
+                    "_metallic": "metallic",
+                    "_normal": "normal",
+                    "_roughness": "roughness",
+                }.items():
+                    if stem.endswith(suffix):
+                        matched_key = key
+                        break
+                textures[matched_key] = tensor
+    if obj_bytes is None:
+        raise ValueError("ZIP archive does not contain an OBJ file.")
+    return ObjZipResult(
+        obj=Types.File3D(source=BytesIO(obj_bytes), file_format="obj"),
+        texture=textures.get("texture"),
+        metallic=textures.get("metallic"),
+        normal=textures.get("normal"),
+        roughness=textures.get("roughness"),
+    )
+
+
 def get_file_from_response(
     response_objs: list[ResultFile3D], file_type: str, raise_if_not_found: bool = True
 ) -> ResultFile3D | None:
@@ -93,6 +162,7 @@ class TencentTextToModelNode(IO.ComfyNode):
                 IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.File3DGLB.Output(display_name="GLB"),
                 IO.File3DOBJ.Output(display_name="OBJ"),
+                IO.Image.Output(display_name="texture_image"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -151,14 +221,14 @@ class TencentTextToModelNode(IO.ComfyNode):
             response_model=To3DProTaskResultResponse,
             status_extractor=lambda r: r.Status,
         )
+        obj_result = await download_and_extract_obj_zip(get_file_from_response(result.ResultFile3Ds, "obj").Url)
         return IO.NodeOutput(
             f"{task_id}.glb",
             await download_url_to_file_3d(
                 get_file_from_response(result.ResultFile3Ds, "glb").Url, "glb", task_id=task_id
             ),
-            await download_url_to_file_3d(
-                get_file_from_response(result.ResultFile3Ds, "obj").Url, "obj", task_id=task_id
-            ),
+            obj_result.obj,
+            obj_result.texture,
         )
 
 
@@ -211,6 +281,10 @@ class TencentImageToModelNode(IO.ComfyNode):
                 IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.File3DGLB.Output(display_name="GLB"),
                 IO.File3DOBJ.Output(display_name="OBJ"),
+                IO.Image.Output(display_name="texture_image"),
+                IO.Image.Output(display_name="optional_metallic"),
+                IO.Image.Output(display_name="optional_normal"),
+                IO.Image.Output(display_name="optional_roughness"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -304,14 +378,17 @@ class TencentImageToModelNode(IO.ComfyNode):
             response_model=To3DProTaskResultResponse,
             status_extractor=lambda r: r.Status,
         )
+        obj_result = await download_and_extract_obj_zip(get_file_from_response(result.ResultFile3Ds, "obj").Url)
         return IO.NodeOutput(
             f"{task_id}.glb",
             await download_url_to_file_3d(
                 get_file_from_response(result.ResultFile3Ds, "glb").Url, "glb", task_id=task_id
             ),
-            await download_url_to_file_3d(
-                get_file_from_response(result.ResultFile3Ds, "obj").Url, "obj", task_id=task_id
-            ),
+            obj_result.obj,
+            obj_result.texture,
+            obj_result.metallic if obj_result.metallic is not None else torch.zeros(1, 1, 1, 3),
+            obj_result.normal if obj_result.normal is not None else torch.zeros(1, 1, 1, 3),
+            obj_result.roughness if obj_result.roughness is not None else torch.zeros(1, 1, 1, 3),
         )
 
 
@@ -431,7 +508,8 @@ class Tencent3DTextureEditNode(IO.ComfyNode):
             ],
             outputs=[
                 IO.File3DGLB.Output(display_name="GLB"),
-                IO.File3DFBX.Output(display_name="FBX"),
+                IO.File3DOBJ.Output(display_name="OBJ"),
+                IO.Image.Output(display_name="texture_image"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -480,7 +558,8 @@ class Tencent3DTextureEditNode(IO.ComfyNode):
         )
         return IO.NodeOutput(
             await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "glb").Url, "glb"),
-            await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "fbx").Url, "fbx"),
+            await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "obj").Url, "obj"),
+            await download_url_to_image_tensor(get_file_from_response(result.ResultFile3Ds, "texture_image").Url),
         )
 
 
@@ -654,7 +733,7 @@ class TencentHunyuan3DExtension(ComfyExtension):
             TencentTextToModelNode,
             TencentImageToModelNode,
             TencentModelTo3DUVNode,
-            # Tencent3DTextureEditNode,
+            Tencent3DTextureEditNode,
             Tencent3DPartNode,
             TencentSmartTopologyNode,
         ]

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"),

compose.yaml

@@ -0,0 +1,49 @@
+# Docker Compose file to run ComfyUI locally using Docker.
+
+services:
+  comfyui:
+    container_name: comfyui
+    build:
+      context: .
+      args:
+        # Declare additional system dependencies for custom nodes
+        APT_EXTRA_PACKAGES:
+        # Set the value of the --extra-index-url flag to select a specific CUDA,
+        # ROCm, or CPU version.
+        PIP_EXTRA_INDEX_URL: "https://download.pytorch.org/whl/cu130"
+
+    ports:
+      - 8188:8188
+
+    # Optional: enable GPU access for hardware acceleration.
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - capabilities: [gpu]
+    volumes:
+      - ./custom_nodes:/comfyui/custom_nodes
+      - ./models:/comfyui/models
+
+      # (Optional) Mount host ComfyUI data directories
+      #
+      #- ./input:/comfyui/input
+      #- ./output:/comfyui/output
+      #- ./temp:/comfyui/temp
+      #- ./user:/comfyui/user
+
+    environment:
+      # Overwrite the container user's UID and GID to match the host's. This
+      # allows files created by ComfyUI to be mounted on the host without
+      # permission issues.
+      UID: 1000
+      GID: 1000
+      # Declare additional Python packages to install. Useful when a custom node
+      # pack does not properly specify all its dependencies or relies on
+      # optional dependencies.
+      PIP_EXTRA_PACKAGES:
+
+    # Optional: Override the default command. In this case, configure ComfyUI to
+    # listen on all network interfaces (which is required when not using
+    # `network_mode=host`.)
+    command: python ./main.py --listen 0.0.0.0

entrypoint.sh

@@ -0,0 +1,63 @@
+#!/bin/sh
+
+# Entrypoint script for the ComfyUI Docker image.
+
+set -e
+
+user="comfyui"
+user_group="$user"
+
+# Allow users to specify a UID and GID matching their own, so files created
+# inside the container retain the same numeric ownership when mounted on the
+# host.
+if [ -n "$UID" ] && [ -n "$GID" ]; then
+    echo "[entrypoint] Setting user UID and GID..."
+    usermod  -u "$UID" "$user" > /dev/null
+    groupmod -g "$GID" "$user_group"
+else
+    echo "[entrypoint] Missing UID or GID environment variables; keeping default values."
+fi
+
+# Changing a user's UID and GID revokes that user's access to files owned by the
+# original UID/GID. To preserve access to runtime data, the ownership of those
+# directories must be updated recursively so that their numeric owner matches
+# the user's new UID and GID.
+echo "[entrypoint] Changing directory ownership..."
+chown -R "$user:$user_group" \
+    /comfyui                 \
+    /home/comfyui
+
+# To use CUDA and other NVIDIA features, regular users must belong to the group
+# that owns the /dev/nvidia* device files -- typically the video group.
+#
+# Known issue: Because these device files are mounted from the host system,
+# there's no guarantee that the device's group ID will match the intended group
+# inside the container. For example, the video group might be mapped to GID 27
+# on the host, which corresponds to the sudo group in the python:3.12 image.
+# This shouldn't cause major problems, and given the lack of a universal
+# standard for system GIDs, there isn't much we can realistically change to
+# address this issue.
+echo "[entrypoint] Adding user to GPU device groups..."
+for dev in /dev/nvidia*; do
+    group=$(ls -ld "$dev" | awk '{print $4}')
+    usermod -aG "$group" "$user"
+done
+
+# Install or update the Python dependencies defined by ComfyUI (or any installed
+# custom node) and also install any user-defined dependencies specified in
+# PIP_EXTRA_PACKAGES.
+echo "[entrypoint] Updating Python dependencies..."
+su -c "
+   pip install                                   \\
+        --no-cache-dir                           \\
+        --disable-pip-version-check              \\
+        --extra-index-url '$PIP_EXTRA_INDEX_URL' \\
+        -r requirements.txt                      \\
+        $(find custom_nodes -mindepth 2 -maxdepth 2 -type f -name requirements.txt -printf "-r '%p' ") \\
+        $PIP_EXTRA_PACKAGES
+" comfyui \
+    || echo "[entrypoint] Failed to install dependencies, starting anyway" >&2
+
+# Run command as comfyui
+echo "[entrypoint] Running command"
+exec su -c "$*" comfyui

manager_requirements.txt

@@ -1 +1 @@
-comfyui_manager==4.1b2
+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==1.41.18
+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>=0.2.10
+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