README.md

metadata

title: Anima V1 CPU Demo
emoji: 🎨
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 5.44.1
app_file: app.py
pinned: false
license: other
models:
  - circlestone-labs/Anima
tags:
  - text-to-image
  - anima
  - zerogpu
  - gradio
  - anime

Anima V1 CPU Demo

A Hugging Face Gradio Space demo for the Anima V1 base release.

This demo intentionally does not load an old Preview Diffusers repo. It uses the current V1 split files from circlestone-labs/Anima:

• split_files/diffusion_models/anima-base-v1.0.safetensors
• split_files/text_encoders/qwen_3_06b_base.safetensors
• split_files/vae/qwen_image_vae.safetensors

It uses DiffSynth-Studio for the actual Anima single-file loader because the upstream Anima repo is published in ComfyUI / Diffusion Single File format rather than as a native Diffusers model_index.json pipeline.

Deploy

1. Create a new Hugging Face Space.
2. Choose Gradio as the SDK.
3. Upload these files.
4. In Settings → Hardware, select CPU.
5. Optional but recommended: enable persistent storage so the model download is cached between restarts.

Startup behavior

The app downloads the required model/tokenizer files from Hugging Face Hub during application startup with huggingface_hub:

• Anima diffusion model, text encoder, and VAE from circlestone-labs/Anima
• Qwen tokenizer files from Qwen/Qwen3-0.6B
• T5 tokenizer files from the public google/t5-v1_1-xxl repo

The app then tries to load the Anima pipeline at module startup. This matches Hugging Face ZeroGPU guidance: GPU functions are decorated with @spaces.GPU, but large models should still be placed on cuda at the root module level rather than lazy-loaded inside the generation function. After a successful startup load, Generate reuses the same in-memory pipeline for that running Space process. If the startup load fails, Generate retries as a fallback and then reuses the loaded pipeline after a successful load.

Why it should not download from ModelScope

DiffSynth-Studio defaults to ModelScope for remote model resolution. This demo sets DIFFSYNTH_DOWNLOAD_SOURCE=huggingface, pre-downloads assets with huggingface_hub, validates the local tokenizer directories, and passes local file paths to ModelConfig(path=...), so generation should not trigger ModelScope downloads.

The T5 tokenizer no longer depends on the gated stabilityai/stable-diffusion-3.5-large repo or any SD3.5 mirror. If you override ANIMA_T5_TOKENIZER_ID to a mirror that does not contain usable tokenizer files, the app falls back to google/t5-v1_1-xxl.

The app stores files under /data/models/anima-v1 when persistent storage is writable, otherwise /tmp/models/anima-v1.

Configuration

Environment variables:

Variable	Default
ANIMA_MODEL_ID	circlestone-labs/Anima
ANIMA_DIFFUSION_FILE	split_files/diffusion_models/anima-base-v1.0.safetensors
ANIMA_TEXT_ENCODER_FILE	split_files/text_encoders/qwen_3_06b_base.safetensors
ANIMA_VAE_FILE	split_files/vae/qwen_image_vae.safetensors
ANIMA_QWEN_TOKENIZER_ID	Qwen/Qwen3-0.6B
ANIMA_T5_TOKENIZER_ID	google/t5-v1_1-xxl
ANIMA_T5_TOKENIZER_SUBFOLDER	unset / empty
ANIMA_T5_TOKENIZER_FALLBACK_ID	google/t5-v1_1-xxl
ANIMA_LOCAL_MODEL_DIR	/data/models/anima-v1, fallback /tmp/models/anima-v1
ANIMA_LOAD_AT_STARTUP	1
ANIMA_VRAM_LIMIT_GB	unset

Why not DiffusionPipeline.from_pretrained(...)?

The upstream V1 repo provides single-file ComfyUI-style weights, not a native Diffusers folder with model_index.json, split component configs, and already-converted component state dicts. A pure Diffusers Space becomes straightforward once a V1 Diffusers conversion exists. Until then, this Space avoids stale preview repos and loads the V1 files directly.

License

The Anima model is under the CircleStone Labs Non-Commercial License and is also subject to NVIDIA's Open Model License Agreement insofar as it applies to derivative models. Review the upstream model card before deploying or sharing outputs.