MiniMax H3 profiling: VAE decode takes 43% of generation time (5.5 min per 15s clip on 4090)
Prestigious_Cat85 · reddit · 2026-08-20
After profiling 75 MiniMax H3 generations, a user identified VAE decoding as the main bottleneck. Generating a 15-second (362-frame) clip on an RTX 4090 takes about 5.5 minutes just for VAE decoding (0.85-1.0 s/frame), not sampling. Attempts like fp8mix VAE (3% gain), merging tasks (no amortization), and Sol-Attn (version conflict) were made. The user asks if a faster VAE exists, if CUDA 13/newer torch helps, and if tiling or temporal chunking can reduce decode time.
More from Multimodal
- PrunaAI's P-Video-2 Pro adds image/video/audio references from $0.01/s — aziz4ai · 2026-10-16
- ElevenLabs blind tests Eleven v4 voice model, wins all 8 languages on expressiveness — JenniferHli · 2026-10-16
- HyperFrames adds Reference Any Video and auto Brand Kits for AI video agents — alifcoder · 2026-10-16
- 'Alien Metamorphosis': AI short made with Grok Imagine + Topaz — creatoroff · 2026-10-16
- Early local test of distilled FLUX.3 Klein: mediocre t2i, editing hopes remain — Ant_6431 · 2026-10-16
- Krea 3 vs Qwen 2.1 vs ZIT: better facial nuance, but unsettlingly realistic outputs — Crazy-Repeat-2006 · 2026-10-16