Anima Model Loader

The moment you download Anima you run into a problem the normal ComfyUI loaders can't touch. Anima doesn't ship as one checkpoint. It's a diffusion transformer on NVIDIA's Cosmos-Predict2 backbone, and Circlestone Labs distributes it as three separate files - the diffusion model, the Qwen3-0.6B text encoder, and the Qwen-Image VAE. The stock "Load Checkpoint" node expects one merged file and gives you nothing. AnimaModelLoader is the part of the Anima LoRA Comparison pack that wires all three up in a single node, so you can stop thinking about it and get to the part where you compare LoRAs.

How it works

It's a thin, honest wrapper over ComfyUI's own loading functions - comfy.sd.load_diffusion_model, load_clip, and comfy.sd.VAE. No exotic dependencies: the pack's pyproject.toml declares zero pip packages beyond ComfyUI itself, so the only thing this node needs is a reasonably current ComfyUI. Three wires go in, three wires come out.

The inputs that matter

The inputs you'll actually set:

• UNET Model (models/diffusion_models) - the Anima diffusion model, e.g. anima_base_v1.0.safetensors.
• CLIP Model (models/text_encoders) - the Qwen3-0.6B encoder from the same release.
• CLIP Type - this is the one to get right. The default is stable_diffusion, and for Anima you want cosmos. It tells the loader which CLIP implementation to build, and the Qwen encoder will not behave correctly under the wrong type.
• VAE Model (models/vae) - the Qwen-Image VAE.
• Weight Precision - default keeps the file's own precision; fp8_e4m3fn and fp8_e5m2 cast the diffusion model to fp8 at load time. That's the lever to pull if you're on a 6GB card, which Anima genuinely runs on - just slower than you'd like.
• CLIP Device - default, or force the text encoder onto cpu if VRAM is tight.

Outputs are MODEL, CLIP, and VAE, which feed the sampler, the CLIP Text Encode node, and the VAE input of the sampler respectively - all standard sockets, so this plugs into any workflow, not just this pack's own nodes.

Troubleshooting

Where people get burned: the CLIP Type dropdown. Miss cosmos and the loader doesn't always throw an error - the code does a getattr(comfy.sd.CLIPType, "COSMOS", ...) with a fallback to stable_diffusion, so on an older ComfyUI that doesn't know the cosmos CLIP type it silently loads the wrong kind of encoder and you get garbled conditioning with no crash to explain it. If text encoding looks wrong, check both the dropdown and that your ComfyUI is current.

The other classic: empty dropdowns. The file lists are read from models/diffusion_models, models/text_encoders, and models/vae, so if a menu is empty the file is in the wrong folder. The tooltips on the fields spell these paths out for you. And if you drop a new model in while ComfyUI is running, restart it - file lists are built at startup.

Install

Install it with ComfyUI Manager (search "Anima LoRA XY") or the usual:

cd ComfyUI/custom_nodes
git clone https://github.com/yunqiankuangyu/comfyui-anima-lora-comparison.git

Then restart ComfyUI. You'll find all four nodes under Anima/LoRA Comparison in the menu, and this loader is the one that starts every comparison workflow. It's not flashy - but for Anima users it's the door.