VAE round-trip test: MiniMax-H3-Image latent compression scores worst, below even SD 1.5
ostrisai · x · 2026-11-12
ostrisai ran a round-trip VAE comparison to explain why H3 Image struggles with detail: its latent space is heavily compressed.
Quality ranking from the test (PSNR/SSIM/LPIPS mean rank):
- FLUX.2 (128x32x32 latent, 16:1) — best across all metrics: 31.56 dB PSNR, 0.9571 SSIM
- FLUX.1 — 28.95 dB
- Wan 2.1 — 26.77 dB
- P-H3-Image — 23.17 dB
- SDXL — 22.45 dB
- SD 1.5 — 22.04 dB
- MiniMax-H3-Image — 21.63 dB, 0.7391 SSIM, worst on every metric, even below SD 1.5
The H3 VAE uses 24x32x32 latent at 32:1 compression — twice FLUX.2's ratio — and facial crops show clear degradation. Takeaway: a model's detail ceiling is largely bounded by its VAE compression.
More from Multimodal
- Izwi: open-source offline voice cloning in Rust, runs on 16GB M1 Pro — zinyando · 2026-11-12
- MiniMax H3 Image generates 3D game scenes in 10s, over 3x faster than GPT Image — HeyToha · 2026-11-12
- Dev releases local world model turning any image into interactive video, runs on RTX 3060 — lucidml_lover · 2026-11-12
- Ex-Meta CV researcher builds AdaL v2, a chat-driven video agent harness for GTM — Scobleizer · 2026-11-12
- MiniMax H3 image gen runs locally on AMD RX 9070, 25-30s per 4MP image — omglionheaded · 2026-11-12
- One reference, one prompt, three image models: side-by-side Nano Banana Pro vs GPT Image 2.6 test — gizakdag · 2026-11-12