mellow_pencil-XL_clear

• Guide v2 (External site): English | Japanese

• Guide v1 (External site): English | Japanese

• Hugging Face

• SDXL_anime_clear_vae

mellow_pencil-XL_clear_v2

mellow_pencil-XL_clear_v2 | mellow_pencil-XL_clear_v1

The previous mellow_pencil-XL_clear_v1 was a model specifically enhanced with intermediate noise to create a sharp, high-definition output.

While this increased contrast and detail, it sometimes resulted in a "burnt" color look or left behind color blotches/unevenness caused by noise artifacts.

In mellow_pencil-XL_clear_v2, I have slightly reduced the intermediate noise to achieve natural color expression and a cleaner texture. As a trade-off, the density of fine background details is slightly lower compared to v1.

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 (v2)

• CFG scale: 3 (2-4)

• Sampler: euler a

• Scheduler: beta

• Steps: 12-20 (around 15)

Quality Tags and Negative Prompts are not always necessary.

Merge Recipe (v2)

• UNET: mellow_pencil-XL-v1.0.0-base

  • input_blocks.0: x0.85

  • output_blocks.6: x1.1

  • out.: x1.05

• CLIP: mellow_pencil-XL-v1.0.0-base

• VAE: SDXL_anime_clear_vae

mellow_pencil-XL_clear_v1

mellow_pencil-XL-v1.0.0-base_clear | mellow_pencil-XL-v1.0.0-base (Original)

Fine Tuned Model of mellow_pencil-XL series, reduce noise for a clearer look and ensures vivid color expression even when the CFG scale is set low.

Recommended Settings (v1)

• CFG scale: 3 (3-5)

• Sampler: euler a

• scheduler: beta

• Steps: 15-25 (around 20)

Quality Tags and Negative Prompts are not always necessary.

Merge Recipe (v1)

• UNET: mellow_pencil-XL-v1.0.0-base

  • time_embed.: x0.90

  • input_blocks.0: x0.90

• CLIP: mellow_pencil-XL-v1.0.0-base

  • clip_l.encorder.layers.4: x0.95

  • clip_g.encorder.layers.13: x0.95

  • clip_g.encorder.layers.31: x0

• VAE: SDXL_anime_clear_vae

Upscaler

your choice, examples:

• RealESRGAN_x4Plus Anime 6B (Commercial use)

• IllustrationJaNai_V1_ESRGAN (No-Commercial use)

Detailer

works well.

Tips

• Color Prompts: English | Japanese

• Color Correct custum nodes in ComfyUI: English | Japanese

• About VAE: English | Japanese

• Illustrious-XL series License: English | Japanese

License

The mellow_pencil-XL_clear model released here is subject to the restrictions of both of the following two licenses:

• Fair AI Public License 1.0-SD

• CreativeML Open RAIL-M

The key points of these two licenses are the requirement to publish the recipe when releasing a derivative model, and furthermore, that Onoma AI, the source of Illustrious-XL 2.0, retains the right to modify the derivative model and its outputs.

We have inquired with Onoma AI regarding the detailed license for derivative models of Illustrious-XL 2.0, but until a response is received, both of the above restrictions shall apply.