vLLM-Omni ships Day-0 RL recipe for MiniMax-H3-Image with 3-term reward
vllm_project · x · 2026-11-10
The vLLM project announced Day-0 RL support for MiniMax-H3-Image in VeRL-omni: a LoRA trained with DiffusionNFT, using FSDP2 for training and vLLM-Omni for rollout, directly on the model's native 6-step distilled sampler. The reward multiplies three terms — OCR score, a layout guard, and an HPSv2.1 quality gate. The training curve shows validation score rising from 0.44 to 0.7 over 50 steps; the recipe landed as a PR in verl-project/verl-omni. The accompanying vLLM-Omni stack supports text-to-image and reference-image editing with multi-GPU VAE tiling, CUDA Graph replay, and a faster image-response path for product scenes, game assets, posters, and UI mockups.
More from Multimodal
- Minimax H3 Context node tutorial: extend AI video generations indefinitely in ComfyUI — Hrmerder · 2026-11-11
- SpaghettiSplats: a strict generalization of 3DGS extending splats to lines and curves — taiyasaki · 2026-11-11
- SpaghettiSplats: closed-form 3D convolution with EWA matches 3DGS-MCMC frame rates — taiyasaki · 2026-11-11
- LiveAvatar upgrades: single-image 1080p full-body avatars with 6 emotions — nikola_mr64990 · 2026-11-11
- Seedream 5.5 vs Magnific One vs GPT Image 2.6 vs Nano Banana 2.1: same-prompt image model showdown — azed_ai · 2026-11-11
- ElevenLabs' Search for the World's Catchiest Ad ends: 11 creators split $100k in prizes — aziz4ai · 2026-11-11