ComfyUI-NVIDIA-CMD

ComfyUI-NVIDIA-CMD

ComfyUI custom nodes that run NVIDIA CMD causal few-step image-to-video on Windows Native.

No WSL. No flash-attn package. Attention is PyTorch SDPA.

日本語 README

This project is unofficial. It is not published, endorsed, or supported by NVIDIA. It is an adapter that calls the official nv-tlabs/cmd checkout. Official CMD code and student weights are licensed under the NVIDIA OneWay Noncommercial License (research and education only). See License.

Built on NVIDIA Cosmos. Version 0.2.0.

Demo

Input image (repo copy: outputs/comfyUI/sampleimage.png; ComfyUI UI name car-red.png):

CMD input image

Workflow	Video	Graph
Short I2V	cmd_i2v_basic.mp4	cmd_i2v_basic
Long video	cmd_long_basic.mp4	cmd_long_basic
Camera control	cmd_camera_control.mp4	cmd_camera_control

Standalone sample: outputs/cmd_i2v.mp4 (93 frames, backend=sdpa).

Why this project

CMD is not a new standalone architecture. It is a distilled causal few-step student of Cosmos-Predict2.5-2B (arXiv:2608.13391). This adapter keeps the official Linux / FlashAttention stack optional and runs the student on Windows Native with Blackwell-tested PyTorch SDPA.

• Windows Native, no WSL required
• RTX 50 / Blackwell: PyTorch SDPA instead of a flash-attn wheel
• Short I2V, long rollout, and camera-control checkpoints as ComfyUI nodes
• Official sources stay in third_party/cmd or CMD_UPSTREAM; they are not vendored

Features

• Nodes: NVIDIACMDModelLoader, NVIDIACMDImageToVideo, NVIDIACMDCameraControl, NVIDIACMDLongVideo, NVIDIACMDSaveVideo
• BF16, 832x480, 16 fps mp4 via NVIDIACMDSaveVideo into ComfyUI output/
• Circular KV cap at local_attn_size so long rollout does not keep every latent frame
• No KSampler compatibility. Negative prompt / CFG are unused in student inference
• No automatic weight download

Status matrix

Verified on the tested environment:

• chunk1_short / cmd_i2v_basic.json
• chunk1_long / cmd_long_basic.json (126 latent frames, about 501 pixel frames, KV capped to 21)
• chunk1_camera / cmd_camera_control.json with examples/identity_camera.npz
• pytest import, node mapping, workflow JSON, camera path, KV cap, runtime-guard restore
• Attention backend sdpa without flash-attn

Not verified in this repository:

• Linux, RTX 40-series, or ComfyUI builds other than the portable stack below
• chunk4_* checkpoints (presets exist; no measured workflow)
• FP8, SageAttention, FlashAttention 2/4 as a required path
• Process-level isolation of official utils / pipeline imports

Tested environment

Measured, not assumed:

• Windows 11 Native
• RTX 5090 32GB, sm_120
• PyTorch 2.9.1+cu130 (the Blackwell build already in ComfyUI; this node does not replace torch)
• ComfyUI Portable

Also exercised with ComfyUI-Win-Blackwell. That stack is tested with, not required. Any ComfyUI install that already has a working Blackwell PyTorch can be used.

This adapter uses the transformers package already present in the ComfyUI environment. Do not install the official CMD requirements.txt into that environment.

Install

Clone into ComfyUI custom_nodes and install adapter dependencies only, with the same Python ComfyUI uses.

cd <ComfyUI>\custom_nodes
git clone https://github.com/hiroki-abe-58/ComfyUI-NVIDIA-CMD ComfyUI-NVIDIA-CMD
cd ComfyUI-NVIDIA-CMD
.\<ComfyUI-python> -m pip install -r requirements.txt
git clone https://github.com/nv-tlabs/cmd.git t