🪞 UMM-Reflection — BAGEL-RL

A unified multimodal model that reflects on its own renders: generate → score → edit → re-render, all in one pass, no external critic. Checkpoint: YijiaFan/UMM-Reflection-BAGEL-RL · Paper · Code

Every run streams the full trajectory below: the first render, then up to 3 reflection rounds where the model writes its score, reasoning and edit instruction verbatim.
10 100
1 8
Resolution

Eval / RL training use 512×512
1 3
0 1

Approximate KV cache between diffusion steps — faster, sanctioned by the repo

Experimental AOT compile of the transformer; first call pays compilation
Example prompts (GenEval-style, instruction voice — the voice the policy was trained on)
Pages:
About the checkpoint

UMM-Reflection-BAGEL-RL is BAGEL-7B-MoT fine-tuned with interleaved Flow-GRPO: 16 sibling trajectories per prompt share one first image, and a single trajectory-level advantage updates both the text (controller) and flow (image) heads. GenEval 0.84, WISE 0.74, OneIG 0.83, T2I-CompBench++ 0.55 (transfer — never seen in training). The first-image accuracy is unchanged from the base model (~0.71); what RL adds is the repair: wrong first images are fixed 64.9% of the time, vs 20.6% before.

Inference settings here mirror scripts/eval/generate_geneval553.py: 512×512, 50 steps, CFG 3.0, controller temperature 0.5, max 3 repair rounds, strict V20 controller protocol.