HOMOCHI XL

HOMOCHI XL is the successor to my MelloMochi SD1.5 model and uses a hybrid tagging system of Florence-2 natural language style prompts and WD14 tags. This gives it better control over composition and physique. It can go between slender Ikemen styles and high-detail Bara or muscular builds without losing anatomical accuracy. With the proper prompting this model can go between 90's retro anime to more 2.5D modern art styles without straying too far from your prompt.

Quick Start Specs

NOTE: Civitai will not show the proper sampler in photos in the gallery created with Restart and will show DPM++ 3M SDE Karras by default

• Primary Sampler: Restart (Highly preferred for the best clean anime lines and detail)

• Alternative Samplers: DPM++ 3M SDE Karras (for a cinematic look) or Euler a (fastest for low spec PCs)

• Sampling Steps: 30 to 35 (up to 50 with Restart)

• CFG Scale: 5.0 to 7.0

• Clip Skip: 2

Supported Resolutions

note that the model was trained primarily on 2:3/3:2 and 9:16/16:9 images though with the right prompting other resolutions can work.

Square

• 1:1 — 1024 x 1024

Portrait (Tall)

• 3:4 — 896 x 1152

• 2:3 — 832 x 1216

• 9:16 — 768 x 1344

Landscape (Wide)

• 4:3 — 1152 x 896

• 3:2 — 1216 x 832

• 16:9 — 1344 x 768

Baseline Negative Prompt

This model responds well to both Florence-2 prompting and WD14 tagging allowing you to use much more detailed negative prompts when needed. Below is an extensive negative geared towards keeping the "Homochi" style I've aimed for with this model:

(an image that was drawn by an amateur artist:1.3), (simple details and incohesive lighting), (the artist has poor art direction and lacks a unique point of view),(the shading and shadows lack depth and dimension),lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, simple background, flat lighting, deformed, poorly drawn, (plastic skin:1.2)

Alternate Negative Prompt

If you find the above isn't working well with your preferred scheduler or is overriding the art style you're aiming for I recommend the more universal negative prompt below:

lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, simple background, flat lighting, deformed, poorly drawn, (plastic skin:1.2)

Workflow

Method 1: Preferred

Use this if you have a decent GPU with 12GB+ VRAM. Upscale the image FIRST, and then inpaint the massive canvas so we don't lose any high-res hand-drawn details.

1. Base Generation (txt2img)

• Sampler: Restart

• Steps: 30 to 35 (35+ YMMV)

• Resolution:

• Square: 1024 x 1024

  Portrait: 832 x 1216

  Portrait (Tall): 768 x 1344

  Landscape: 1216 x 832

  Landscape (Wide): 1344 x 768

2. The High-Res Pass (Ultimate SD Upscale): Before fixing faces or hands, bring the whole image up to its final size. Send your base image to img2img.

• Denoising Strength: 0.25 to 0.45

• Script: Ultimate SD upscale

• Target size type: Scale from image size

• Scale: 2

• Upscaler: 4x_NMKD-UltraYandere_300k (Highly recommended to keep anime lines clean)

• Type: Chess

• Tile width/height: 1024x1024

• Adjust padding if the final image has seams

3. Final Inpainting: Now send it to the Inpaint tab to fix the eyes, hands, or background details.

• Inpaint Area: Only Masked

• Resize to: 1024x1024 (Vital: This gives the AI a high-res window into your 4K image)

• Only Masked Padding: 32 to 64

• Denoising Strength: 0.45 to 0.60

• Prompt: Just prompt the part you are fixing (e.g., "close up of beautiful anime eyes, blue")

Method 2: The Low VRAM Pipeline (For Older PCs)

Use this if Method 1 gives you an "Out of Memory" error. Loading a 4K image into the inpaint canvas can crash 8GB cards. Here, we fix the anatomy first at base resolution, then run a highly optimized upscale.

1. Base Generation

• Sampler: Euler a (lightest on hardware) or DPM 3M SDE

• Keep base generation exactly the same (832x1216 or 1024x1024 ex).

2. Inpaint First Send the base image straight to the Inpaint tab to fix any weird anatomy before making it bigger.

• Inpaint Area: Only Masked

• Resize to: 1024x1024

• Denoising Strength: 0.40 to 0.60

3. Optimized Upscale Once the anatomy looks right, send the fixed image to img2img for the final upscale. We drop the tile size here to save your VRAM.

• Denoising Strength: 0.25 (Keep this low so the upscaler doesn't erase your inpainting fixes)

• Script: Ultimate SD upscale

• Scale: 1.5 or 2

• Upscaler: 4x_NMKD-UltraYandere_300k

• Type: Chess

• Tile width/height: 512x512 (Lowering this to 512 prevents crashes on low-end cards)