MicksHF's Qwen Anime FUSAL Brings Multi-Character Design Sheets to Open Image AI
A 400MB anime-style LoRA for Qwen Image 2.1 brings curated aesthetics, character sheet conditioning, and open weights without a gatekeeper.
Takeaways
- New Qwen-Image-2.1-Anime-Fusal LoRA adapts Qwen Image 2.1 to anime in a ~400 MiB file
- Trained on roughly 50,000 manually curated text-to-image and reference-to-image records
- Supports multi-character and multi-reference generation from IP character design sheets
- Ships original-base rank-80 adapter plus a rank-128 Turbo E6 preview for the Turbo base
- Uses structured bracket prompts with a
fusal style.trigger, 25 steps, CFG 1.0, strength 0.7-1.0 - Released under the Qwen Research License for non-commercial use; try the public Playground Space
MicksHF has released Qwen Anime FUSAL, an anime-focused LoRA for the official Qwen Image 2.1 checkpoints. The main adapter is about 400 MiB and targets text-to-image generation, character-design-sheet references, multi-character composition, and NSFW illustration. A LoRA stores a compact set of weight updates that attaches to an existing model, allowing developers to add the specialization without deploying another full image-generation stack.
The public repository requires no access approval and links to a Playground Space for browser-based testing. The files use the Qwen Research License, which limits use to non-commercial research unless the user obtains separate authorization from Qwen.
Two adapters, strict pairing
| Release | Required checkpoint | Rank | Approximate size | Status |
|---|---|---|---|---|
| FUSAL v2 | Qwen Image 2.1 Base | 80 | 400 MiB | Primary release |
| FUSAL Turbo E6 | Qwen Image 2.1 Turbo | 128 | 640 MiB | Preview |
Each adapter must remain paired with its designated checkpoint. Loading the Base adapter on Turbo, loading the Turbo adapter on Base, or stacking both adapters can produce incorrect results because the underlying model weights and sampling behavior differ. The Turbo E6 file remains under active iteration, so later releases may change its output.
The repository also provides a matching BF16 Heretic text encoder. A text encoder converts the prompt into embeddings that condition the image model, while BF16 is a 16-bit numerical format that reduces memory use. Training cached prompt and reference embeddings with this encoder, so using another encoder can shift prompt interpretation. The official VAE, processor, and tokenizer remain the recommended supporting components.
What FUSAL changes
FUSAL combines the base model’s semantic capabilities with manual aesthetic curation. According to the model card, training used roughly 50,000 text-to-image and reference-conditioned records selected from a substantially larger candidate pool. The adaptation concentrates on anime composition, anatomy, character interaction, and stylistic consistency.