ostris trains H3 image VAE overnight with face-identity and OCR losses, benchmark vs MiniMax
ostrisai · x · 2026-11-13
Indie developer ostris shared an overnight fine-tune of the H3 image decoder, adding facial identity and OCR losses on top of lpips, fft, and l1 losses, without a critic loss.
The attached benchmark shows his P-H3-Image VAE (32:1 compression, 26.2M params) still ranking first on MSE/PSNR/SSIM/LPIPS, ahead of Ostris H3 Image VAE and MiniMax-H3-Image — though he notes it's a significantly larger decoder. Training continues.
More from Multimodal
- Visual AI agents should save discovered pipelines as reusable programs, argues dev — prateekj · 2026-11-13
- One-prompt video editing: Fable 5.2 cuts an entire clip from scratch — BLCNYY · 2026-11-13
- Fei-Fei Li's team launches Atlas: agents build coarse 3D, model renders pixels — drfeifei · 2026-11-13
- Script in, film out: Claude Fable 5.2 makes a full trailer via Higgsfield MCP — techhalla · 2026-11-13
- PortalCam Gaussian splatting scan of an aviation museum, viewable in SuperSplat — willeastcott · 2026-11-13
- Synthesia lands on inaugural AI 100 list, launches AI video agent Syren — synthesiaIO · 2026-11-13