Cagliostro's Animagine XL 4.0 Fixes Anatomy and Color After 800K Downloads
Cagliostro Lab's open-source anime SDXL finetune ships with 8.4M training images, a tag-based prompt grammar, and a refined Opt variant for better anatomy and color.
- Cagliostro Lab released Animagine XL 4.0, a full SDXL retrain for anime generation.
- Trained on 8.4M images for ~2650 GPU hours on 7x H100 80GB.
- Opt variant improves anatomy, color saturation, noise, and output stability.
- Uses tag-based prompts with quality, score, temporal, and rating special tokens.
- Recommended settings: CFG 5, 28 steps, Euler Ancestral, 832x1216 portrait.
- Open RAIL++-M license permits commercial use; weights are fp16 safetensors.
Cagliostro Research Lab has released Animagine XL 4.0, an anime-focused full-model fine-tune of Stable Diffusion XL 1.0, followed by an optimized Opt checkpoint. The newer variant targets inconsistent anatomy, residual noise, muted saturation, and color drift while retaining the same SDXL architecture and tag-based prompting system.
At publication, the Hugging Face repository had recorded more than 800,000 downloads. That metric counts file requests and may include repeat downloads from the same user. Existing SDXL front ends can load the fp16 SafeTensors weights, although prompt behavior and adapter compatibility differ from vanilla SDXL and earlier Animagine releases.
Two checkpoints, different jobs
Version 4.0 started from SDXL 1.0 and updated the UNet and text encoders on 8.4 million anime-style images. Updating the core weights produces larger behavioral changes than adding a LoRA, including stronger character knowledge and closer adherence to Danbooru-style tags.
The Opt checkpoint keeps the same architecture while refining the initial release. Cagliostro reports more consistent generations, improved body proportions, fewer artifacts, richer saturation, and more accurate colors. Opt is the practical choice for new integrations; the original checkpoint remains useful when reproducing images or workflows built against the first 4.0 release.
| Use case | Recommended choice | Reason |
|---|---|---|
| New anime generation workflow | Opt | Includes the anatomy, noise, and color refinements. |
| Reproducing existing 4.0 output | Base 4.0 | Preserves the initial checkpoint’s behavior. |
| Matching a specific illustrator | Opt with a compatible style LoRA | The base training emphasizes character identity more than exact style consistency. |
Inside a 2,650 GPU-hour run
The model card reports the following training configuration:
| Component | Configuration |
|---|---|
| Base model | Stable Diffusion XL 1.0 |
| Training data | 8,401,464 anime-style images from multiple sources |
| Dataset cutoff | January 7, 2025 |
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
