Xiaohongshu's DeCoRL decomposes consistency rewards to fix multi-image composition
小红书技术REDtech · wechat · 2026-11-09
Xiaohongshu's AIGC team released DeCoRL, a reinforcement learning framework tackling identity confusion, attribute drift, and missing objects in multi-image composition (2+ reference images).
Instead of scoring whole images, DeCoRL decomposes consistency into per-object judgments: it matches objects across reference and generated images, uses a specially trained single-object consistency reward model to check identity and appearance for each, then aggregates these into a multi-object reward — with penalties for missing objects — to drive RL post-training, alongside a holistic quality score.
Results: on their new DeCo-Bench (200 real-photo cases, 2–6 reference images), the overall score rose from 7.05 to 8.16, and from 6.15 to 7.61 in the hardest 6-image setting. On public MICo-Bench, the score jumped from 37.58 to 54.03 — the highest among models listed in the paper, with notable gains in person composition and decomposition-recombination tasks. Use cases include group photos from multiple sources, product-scene combinations, and outfit recomposition.
More from Multimodal
- Drawing every frame in JavaScript: Fable 5.2 code-generated video has no nudity filters — mickeyxfriedman · 2026-11-11
- pDMD LoRA brings 2-step realtime editing to MiniMax's 33B image model at 16MP — _akhaliq · 2026-11-11
- Gemini 4 Carbon rebuilds 220m of downtown Los Gatos in Blender with near-zero manual modeling — bilawalsidhu · 2026-11-11
- User asks Fable 5.2 to paint the most complex image it can — a da Vinci-style tour de force — imjustnewatai · 2026-11-11
- "Midnight Melody": emotional street-singer AI video prompt, 9:16 VFX — propmt07 · 2026-11-11
- First AI opera film showcases hybrid filmmaking: coding meets AI generation, editing and VFX — teortaxesTex · 2026-11-11