---
title: Hassaku-XL
canonical_url: "https://www.modelscope.cn/models/ModelE/Hassaku-XL"
md_url: "https://www.modelscope.cn/models/ModelE/Hassaku-XL.md"
repository: ModelE/Hassaku-XL
last_updated: 2025-10-02
license: other
pipeline_tag: text-to-image-synthesis
tasks:
  - text-to-image-synthesis
base_model:
  - AI-ModelScope/stable-diffusion-xl-base-1.0
base_model_relation: finetune
parameters: 3.5B
tensor_type:
  - F16
library_name:
  - lora
  - safetensors
  - pytorch
frameworks:
  - Pytorch
language:
  - en
supports_inference: txt2img
downloads: 4867
stars: 54
tags:
  - Checkpoint
  - text-to-image
---

# Hassaku-XL

> Hassaku-XL - ModelE 在 ModelScope 开源的模型。Hassaku aims to a model with a distictive anime look, capable to do great sfw and nsfw art.V1 use illustrious-xl as base, versions from 0.2M up are trainined with quality prompts like masterpiece or worst quality.…

ModelE/Hassaku-XL 是 ModelScope 魔搭社区上的 3.5B 参数text-to-image-synthesis模型，采用 other 许可，基于 AI-ModelScope/stable-diffusion-xl-base-1.0 构建，并支持在线推理（txt2img）。

- **Repository**: ModelE/Hassaku-XL
- **License**: other
- **Tasks**: text-to-image-synthesis
- **Parameters**: 3.5B
- **Base model**: AI-ModelScope/stable-diffusion-xl-base-1.0
- **Online inference**: txt2img
- **Tags**: Checkpoint, text-to-image
- **Downloads**: 4867
- **Stars**: 54
- **Last updated**: 2025-10-02

Source: https://www.modelscope.cn/models/ModelE/Hassaku-XL

---

# Hassaku - Anime-Optimized XL Model

A high-fidelity anime-style diffusion model designed for clean, consistent, and expressive character generation.

Hosted version and API access available at:
https://sinkin.ai/m/JWknjgr

---

## About the Model

Hassaku is a fine-tuned Stable Diffusion XL model trained specifically for anime-style artwork. It emphasizes visual clarity, minimal noise, and natural composition, with training focused on high-quality illustrations free from common artifacts.

This model performs best when prompts align with its training data distribution. Avoid introducing elements not present during training to achieve optimal results.

---

## Key Features

- **Clean outputs**: No floating text, logos, speech bubbles, or watermarks in training data
- **Simplified prompting**: Designed for straightforward and predictable tag usage
- **Consistent style**: Bright, sharp anime aesthetic with strong facial and pose coherence

Use the negative prompt `signature` if artifacts such as signatures or text appear.

Do not use metadata tags like `highres`, `masterpiece`, or franchise-specific tags such as `re:zero kara hajimeru isekai seikatsu`. These were excluded from training and have no effect.

---

## Prompt Structure

Use the following order for best results:

```
[number of people], [character name], [pose/action], [composition], [optional style/lighting]
```

Example:
```
1girl, rem (re:zero), standing, upper body, soft lighting
```

Incorrect usage:
```
masterpiece, best quality, 1girl, rem, highres, standing
```

Avoid redundant quality tags and unstructured ordering.

---

## LoRA Compatibility

NoobAIs LoRAs are generally more compatible and perform better with Hassaku than Illustrious LoRAs.

Support for other fine-tuning methods may vary. Test carefully with your target style.

---

## Recommended Resolutions for SDXL

These aspect ratios are optimized for balanced generation quality:

- 1536 × 640
- 1344 × 768
- 1216 × 832
- 1152 × 896
- 1024 × 1024
- 896 × 1152
- **832 × 1216** *(most recommended)*
- 768 × 1344
- 640 × 1536

Generating outside these resolutions may reduce quality or cause composition issues.

---

## Usage Guidance

Review the example images to understand the intended aesthetic and prompt alignment.

The model expects clear, concise prompts with minimal extra modifiers. Over-prompting or including irrelevant tags may degrade output.

Always include relevant negative prompts such as:

```
signature, text, logo, bubble, watermark, lowres, bad anatomy, extra limbs
```

---

## Support and Community

For help, discussion, and updates related to Hassaku and other anime models, visit the associated Discord channel.

Direct links and invites can be found via the maintainer's profile.

API documentation and real-time inference available at:
https://sinkin.ai/m/JWknjgr

---

## Version Information

- **Below v1**: Based on ANIMAGINE XL 3.0
- **v1**: Merge of Illustrious-XL and WAI-NSFW-Illustrious-SDXL with additional training
- **v2**: Standalone training based on v1; includes newer and missing characters
- **v3**: Merge of Illustrious-XL and v2, improving stability and character coverage

All versions prioritize usability and consistent behavior under simple prompt conditions.

---

## Acknowledgements

Special thanks to supporters:
- Riyu
- SETI
- Kodokuna

Your support enables ongoing testing and development.

---

## License

This model is released under the **Fair AI Public License 1.0-SD**.

Full license terms are available in the [LICENSE](LICENSE) file.

Key conditions:
- Free to use, modify, and share
- Cannot be used to train models that are commercially restricted or closed
- Must provide appropriate attribution

See the license document for complete details.