Fizgig 9.7.0: AdaFace over ArcFace pushes H3-Image LoRA likeness to new highs
shootthesound · reddit · 2026-11-12
After 12 hours of testing, Fizgig v9.7.0 ships with top likeness scores for character LoRA training on MiniMax H3-Image. What's new:
- Retuned H3 Image Standard preset: a new Likeness optimization schedule (shifting which noise levels training focuses on over time), Automagic v4 (fixing over-steering), 0.25 MP and 30 epochs — face shape comes through earliest, likeness peaks highest, model quality holds.
- AdaFace ViT replaces ArcFace for likeness scoring: AdaFace picked the same best epochs the author chose by eye every time; ArcFace often didn't.
- "Your photos" band on the likeness chart: your own training photos are scored against the same references, showing what "as good as a real photo" means for your dataset.
Usage: pick the MiniMax H3 Image Standard preset (0.25 MP) and train with defaults. The author also corrects an earlier move to 0.5 MP — the schedule and optimizer were what limited likeness, and 0.25 MP is faster with less VRAM. On the hardest dataset, the best previews beat 43 of the person's own 45 photos on likeness.
More from Multimodal
- Fable 5.2 compresses 40 years of computer graphics evolution into a 29-second animation — FinanceYF5 · 2026-11-13
- Ant's Ling-3.2-flash demo turns meal photos into illustrated food journals — pipecanon · 2026-11-13
- Minimax-generated video of Jensen Huang handing NVIDIA to Tim Cook goes viral — Riddy_93 · 2026-11-13
- A 'very convincing' grape arbor photo shows how real AI images now look — misovalko · 2026-11-13
- Krea demos 'Agentic Editing' to SF growth crowd, calling it Cursor moment for creative work — angrypenguinPNG · 2026-11-13
- Manga ink backgrounds LoRA for NoobAI V-Pred delivers solid blacks and real linework — Historical_Song1864 · 2026-11-13