Jolly Cosmos3 Model Loader

Every Cosmos3 workflow in this pack starts here. JollyCosmos3ModelLoader is the node that loads NVIDIA's Cosmos3OmniPipeline from diffusers and hands it to the generation nodes as a COSMOS3_PIPE. No pipe, no output - the other four nodes just sit there waiting for it.

Why a dedicated loader at all? Because the ComfyUI stock Cosmos nodes speak Cosmos 1/1.5 (T5 cross-attention, EDM sampling), and Cosmos3-Nano speaks something else entirely: Qwen2 joint-token sequences with unified 3D mRoPE position embeddings and flow matching. This loader wires up the official inference path from the nvidia/Cosmos3-Nano model card, including the part people usually get wrong - the scheduler. It swaps in UniPCMultistepScheduler(flow_shift=10.0), matching training rather than guessing. (Community recipes elsewhere run flow_shift=5.0; this pack follows the model card, which is the defensible call.)

The inputs

• model_name - defaults to cosmos3_nano. You can also pass a local path, a folder under ComfyUI/models/diffusers/, or an HF repo ID like nvidia/Cosmos3-Nano, and the loader will hunt for a model_index.json in models/diffusers/, models/Cosmos3/, or the HF cache.
• precision - bf16 (default) or fp16. bf16 is the safer pick on Ada and newer; fp16 shaves a little VRAM if you're scraping by.
• device - auto is fine for almost everyone. The cuda:0–cuda:3 options exist for multi-GPU boxes.
• disable_guardrails - default true, and the tooltip says it plainly: "Skip Cosmos safety guardrails." Note this only takes effect if the cosmos_guardrail package is actually installed - if it isn't, the safety checker is off regardless.

Output is a single cosmos3_pipe object, wired into JollyCosmos3TextToImage, JollyCosmos3TextToVideo, or JollyCosmos3ImageToVideo. That object carries the whole pipeline - VAE, transformer, tokenizer, scheduler, and the sound tokenizer when the checkpoint has one (that's how the gen nodes know whether generate_sound will do anything).

Installing it - the gotcha

This is a HuggingFace-diffusers loader, not a ComfyUI-native safetensors loader. It loads a whole diffusers model directory, and it requires diffusers from git main:

cd ComfyUI/custom_nodes
git clone https://github.com/JOLLYKRONK/jolly-cosmos3-comfyuinodes
pip install -U 'diffusers @ git+https://github.com/huggingface/diffusers.git'
huggingface-cli download nvidia/Cosmos3-Nano --local-dir ComfyUI/models/diffusers/cosmos3_nano

Restart ComfyUI and the nodes appear under the Cosmos3 category. If you load and get an ImportError saying your diffusers has no Cosmos3OmniPipeline, that's the git-main requirement biting - upgrade and restart. If you get a FileNotFoundError, the error message tells you exactly which directories it looked in, which almost always means the model download landed somewhere unexpected.

The honest hardware talk

Cosmos3-Nano is 16B parameters (8B active, a Mixture-of-Transformers), around 32GB in BF16. Community consensus is that a 5090 is "basically at the limits" of running it properly and the RTX 6000 Pro class is what NVIDIA architected it for. Don't buy this pack expecting it on a 12GB card. The loader does what it can - patches the VAE decode to clear CUDA cache before each call to dodge fragmentation OOMs - but it can't conjure VRAM.

If you're here, you want Cosmos3's world-model behavior (physics, motion, robotics-grade scene understanding) rather than prettier frames - for pure image quality the ecosystem's image models leave it behind, and it knows it.