Replacing 32B with 4B/8B encoders in MiniMax H3 optimization tests
Fit_Ad7343 · reddit · 2026-08-15
The author released v3 of MiniMax H3 text encoder optimization, using smaller Qwen3-VL models (4B/8B) with learned projection matrices to replace the original 32B encoder while keeping the DiT untouched. Results show that 4B/8B versions closely track the 32B on explicit instructions (pose, clothing, action), achieving a cosine similarity of 0.9449 (8B) and reducing VRAM usage from 15.7GB to 5GB. The limitation is that details not encoded by the smaller model cannot be recovered, causing variations in unspecified background elements.
More from Multimodal
- FrankenSim builds a physics-based virtual orchestra, piano and drum demos live — doodlestein · 2026-10-16
- vLLM officially supports Muse Spark Open: multimodal, 1M context, agentic — vllm_project · 2026-10-16
- InfoVis: 9B unified model tops open-source slide generation and editing, trained on 13M records — Jiahao Wang · 2026-10-16
- RepForcing: feature-space supervision improves few-step autoregressive video distillation — zju · 2026-10-16
- SCOPE: semantic span-aware GRPO objective improves T2I prompt rewriting — university-of-rochester · 2026-10-16
- Orbia benchmark asks: do generated videos form a consistent 3D world? — tencent · 2026-10-16