DiffSynth: Anima Loader

Anima is the anime base that made people finally uninstall an Illustrious checkpoint: a 2B diffusion transformer from CircleStone Labs built on NVIDIA's Cosmos-Predict2, with a Qwen3-0.6B text encoder instead of CLIP. That swap is the whole point - Illustrious and NoobAI trained their encoders so hard on Danbooru tags that general world knowledge got overwritten, so tags and natural English fought each other. Anima's never went through that. Tags, prose, or both in one prompt, and multi-character scenes hold together.

This Loader is how you bring it up under DiffSynth rather than through the native ComfyUI path. Same file, different plumbing: you get the four-stage VRAM offload and the pack's quantization options, and you pay for it with a graph that needs five nodes before you see a prompt box.

The sockets that matter

Required: model_configs (the merged list from DiffSynth: Merge ModelConfigs), torch_dtype (bfloat16, float16, float32), device (cuda/cpu).

Optional, and genuinely required in practice for Anima:

• tokenizer_config - a single ModelConfig for the Qwen3 tokenizer directory.
• tokenizer_t5xxl_config - a second tokenizer ModelConfig. Anima's text stack wants a T5 tokenizer side by side with the Qwen one, so you're loading two tokenizer directories even though there's one visible text encoder.
• vram_limit - wire a VRAM Limit node here (cuda, 0.5GB buffer in the templates).

One output: pipe → DiffSynth: Anima Inference.

The shipped template declares it all explicitly, and it's worth copying rather than guessing: circlestone-labs/Anima → split_files/diffusion_models/anima-preview.safetensors, split_files/text_encoders/qwen_3_06b_base.safetensors, split_files/vae/qwen_image_vae.safetensors - note the split_files/ prefix, the repo isn't laid out like a normal diffusers folder. Then Qwen/Qwen3-0.6B with pattern ./ into tokenizer_config, and stabilityai/stable-diffusion-3.5-large with tokenizer_3/ into tokenizer_t5xxl_config. If you point the loaders at the current base v1.0 release rather than the preview, keep the same three-file shape and just change the filenames.

Install

# ComfyUI Manager: search "DiffSynth-ComfyUI" → Install → restart

pip install git+https://github.com/modelscope/DiffSynth-Studio.git
cd ComfyUI/custom_nodes
git clone https://github.com/modelscope/DiffSynth-ComfyUI.git
cd .. && python main.py

ModelConfig nodes download on execute, from ModelScope unless you flip download_source to huggingface. CircleStone's license is non-commercial on the weights - selling your renders is fine, hosting it as a service isn't.

Where people get burned

Anima's VRAM floor is about 6GB and it's slow for a 2B model - roughly 10 seconds on a 5090, over a minute on a 4060, and one user reported 588 seconds for a 1536×1536 render on a 6GB RTX 2060. So build the VRAM Config node with CPU offload on from the start; that's the difference between "usable on 6GB" and an out-of-memory crash. If the node fails with a tokenizer error, you skipped tokenizer_t5xxl_config. And if the whole DiffSynth category is missing from your node menu, DiffSynth-Studio isn't installed - the pack imports it at load time and fails as a unit.