Cosmos模型配置器

Cosmos is NVIDIA's family of "world foundation models" - built for Physical AI, meaning robotics and simulation - and the original Cosmos 1.0 diffusion model was the first one to hit the open-weights scene. It never got much adoption here as a generation model, but if you're fine-tuning a Cosmos 1.0 text2world checkpoint for your own use, this is the node that wires its files into Diffusion-Pipe.

What it does

Exactly what the name says: CosmosModelNode takes full paths to the three model components and hands you a model_path config for GeneralConfig. The inputs are all required:

• transformer_path - the main transformer weights, e.g. cosmos-1.0-diffusion-7b-text2world.pt. Note it's a .pt (PyTorch) file, not a safetensors - NVIDIA's Cosmos release format.
• vae_path - the VAE file. Required, full path.
• text_encoder_path - the text encoder file. Required, full path.

No optional knobs here. It's the plainest model node in the pack: three paths in, one config out. The output dict carries a type: "cosmos" marker that tells diffusion-pipe which model class to build, plus your three paths.

Where it sits in the workflow

The wiring is the standard pack pipeline: CosmosModelNode → GeneralConfig.model_config → Train. You'll pair it with an AdapterConfigNode (set to lora - the README's table shows Cosmos is LoRA-only, no full fine-tune), a dataset config, and an optimizer. NVIDIA's Cosmos checkpoints are big, so don't underestimate the disk and VRAM picture; block swapping (blocks_to_swap on GeneralConfig) is your friend if you're near the edge.

Installing the pack

Same pack, same story - Linux/WSL2 only, submodules required:

cd ComfyUI/custom_nodes/
git clone --recurse-submodules https://github.com/TianDongL/Diffusion_pipe_in_ComfyUI.git
git submodule update    # mandatory - training silently fails without it
pip install -r Diffusion_pipe_in_ComfyUI/requirements.txt

Restart, import example_workflows/DiffusionPipeInComfyUI.json, and read the notice notes in the workflow before you queue anything - the README is explicit that the workflow's built-in prompts carry important dataset-building instructions.

Common issues

The usual suspects apply: missing paths (the node returns an error dict if any of the three is empty, so check the console output when the input turns red), and the pack-wide WSL2 path convention - Windows drive-letter paths (Z:/...) instead of /mnt/z/.... Also, be careful you're grabbing the Cosmos 1.0 diffusion checkpoint and not one of the later predict2 models; this node is specifically for the original 7B text2world. If you have a Cosmos-Predict2 file, use that pack's dedicated CosmosPredict2ModelNode instead - the T5 and VAE requirements are different and mixing them up is a guaranteed bad time.