H3 Image's over-compressed VAE blamed for detail loss as finetune shows early gains

ostrisai · x · 2026-11-12

AI image developer ostris argues H3 Image's weak detail stems from its highly compressed latent space (24x32x32, 32:1 ratio) — its VAE round-trip scores worse than even the SD 1.5 VAE on every metric. He repurposed an existing Primal VAE training script to finetune the H3 Image VAE. After a few hours, all metrics improve: his checkpoint hits PSNR 22.25dB / SSIM 0.7560, beating both the original and MiniMax's H3-Image VAE (PSNR 21.63dB). He plans to let it run for days to see how far it can go.

Related event: Developer Finds MiniMax H3 Image VAE Over-Compression Causes Detail Loss(4 posts)→

Original post →

More from Multimodal

Multimodal channel →