Use CFG 5 for perfect saturation/sharpening IMO
Try "worst quality, low quality, lowres, blurry, pixelated, jpeg artifacts, oversaturated, watermark, signature, unfinished, sketch" negative tags sometimes having less can result in better images or even "worst quality, low quality, lowres, blurry, pixelated, jpeg artifacts, mosaic blur, watermark, signature, unfinished, sketch, bad hands, distorted eyes, bad fingers"
Now trying “worst quality, low quality, lowres, jpeg artifacts, cutoff, grainy, distorted, watermark, error, unfinished, sketch, draft, deformed, poorly drawn face, poorly drawn eyes, poorly drawn hands, fused fingers, bad hands, deformed hands, bad fingers, distorted eyes, censored, mosaic censoring, futanari, monochrome” you can remove futanari if you like it
Or even just "worst quality, low quality, lowres, bad, error" and even better "worst quality, bad quality" can be better as too much negative tags can confuse the ai
Description
1. amanatsullustrious_v2.safetensors
2. hassakuXLIllustrious_v32.safetensors
3. prefectIllustriousXL_40.safetensors
4. oneObsession_v18.safetensors
Merges performed:
- Merged amanatsullustrious_v2.safetensors + hassakuXLIllustrious_v32.safetensors at 85/85.
- Merged oneObsession_v18.safetensors + prefectIllustriousXL_40.safetensors at 85/85.
- Merged the two intermediate results at 85/85, favoring the amanat/hassaku branch (so the final blend leans toward amanat + hassaku v32).
LoRAs:
- Applied StyleArtists\Enhance\Z_S_P_748cm... at 15% strength (model and CLIP).
- Applied StyleArtists\S_P_Momofuki_Rio_Illustrious... at 20% strength (model and CLIP).
Final blend and output:
- Merged the base blend with the LoRA outputs at 50/50 (final merge ratio 0.50).
- Used purecolorVAE_v10.safetensors as the VAE for finalizing colors.
- Saved checkpoint prefix: checkpoints/ComfyUI.
Notes:
- Goal was improved consistency on clothes; final model relies more on amanatsullustrious_v2 and hassakuXLIllustrious_v32.
- I avoided bringing in the lineart-heavy behavior from hassakuXLIllustrious_v13StyleB (not included in this merge).
- Did a lot of tests/finetunes and multiple saves during development, so some recorded ratios/labels might be slightly imprecise.