Final version:
Better hands, faces, eyes.
Reduced NSFW content. LCM with simple, normal, or beta settings. Steps: 6โ12. CFG: 1โ1.4. Clip skip 2. Can generate at 832ร1216 or 1024ร1536.
My custom-merged model produces exceptional results when combined with [LoRAs]. Give it a try!
V1-4 pairs exceptionally well with Simulacrum CLIP-L and CLIP-G.
๐๏ธ V3 Prompt Tips
# POSITIVE PROMPT STARTER:
"masterpiece, Best quality, [your description here]"
# NEGATIVE PROMPT STARTER:
"low quality, worst quality, [unwanted elements]"## ๐ Version Recommendations
- V2: Best for photorealistic outputs
- V1: more like V2 than V3, still great for person loras
- V3: Optimized for artistic quality
## โ๏ธ Recommended Settings (V1-V2)
```yaml
Samplers:
- DPM++ 2M
- DPM++ 2M SDE
Sampling Steps: 30-50
Clip Skip: 1-2
CFG Scale: 4-6
๐ง Optional Enhancements
Use ADetailer for refined faces/hands
Enable Hires.fix for upscaling (recommended: 1.5-2x scale)
Description
Prompt tags: Positive masterpiece, high quality
Negative: low quality, worst quality,
FAQ
Comments (2)
v2 is the best so far. v1 weaker prompt adherence, v3 totally unrealistic (anime proportions, too big eyes etc.); v2 strikes the right balance.
V3 seems like a step backwards compared to V2. Also, it only seems to work with clip skip2 for me which I think accounts for some of the weird results. Not sure what you were after. Still like V2 though.










