anima_pencil-XL_clear

• Guide (External site): English | Japanese

• Hugging Face

Compare to Animagine-XL_3.0 (Original)

Animagine-XL_3.0_clear | Animagine-XL_3.0 (Original)

A fine-tuned version of Animagine-XL_3.0 model, designed to produce clearer and more vibrant outputs.

Compared to the original, this model generates illustrations with higher contrast, richer colors, and more refined details.

INT8_ConvRot_HQ (light weight)

INT8_ConvRot_HQ is a lightweight version. The CLIP and VAE remain unchanged, while only the intermediate layers of the UNET have been converted to INT8_ConvRot format.

Image quality is slightly lower than the original v2, but it is particularly well-suited for fast generation in low-VRAM environments (4-6 GB).

INT8_ConvRot is currently supported in ComfyUI and Stable Diffusion webUI Forge NEO.

Recommended Settings

General

• CFG scale: 4-6

• Sampler: DPM++ SDE, Euler A

• scheduler: beta

• Steps: 15-25 (around 20)

Upscaler

your choice, examples:

• RealESRGAN_x4Plus Anime 6B (Commercial use)

• IllustrationJaNai_V1_ESRGAN (No-Commercial use)

Detailer

works well.

Merge Recipe

• UNET: Animagine-XL_3.0

  • time_embed.: x0.95

  • input_blocks.0: x0.92

• CLIP: Animagine-XL_3.0

  • clip_l.embeddings: x0.5

  • clip_l.encorder_layers.1: x0

  • clip_g.encorder_layers.31: x0

• VAE: SDXL_clear_vae

Tips (External site)

• Steps and Scheduler: English | Japanese

• Upscalers: English | Japanese

• Negative Prompt: English | Japanese

• Improving Prompt Encoding in SDXL: English | Japanese

License

• Fair AI Public License 1.0-SD

• License Official Page