Walt Stephen

Model · Any to any

UMM-Reflection-BAGEL-RL

Walt Stephen

The final model of UMM-Reflection (Learning Native Reflection in Unified Models). Starting from it is trained with multi-round Flow-GRPO. For each prompt, it samples 16 complete reflection trajectories from the same first image and optimizes the text (controller) and flow (image) heads jointly against a graded six-family GenEval reward. Local evaluation with the same prompts, resolution, and scorer for every row. GenEval uses the 553 official prompts with up to three repair rounds. On GenEval-553, 64.9% of images that are wrong after the first render end up correct after reflection, against 20.6% for the SFT model. The layout is the same as the base BAGEL repository. ema.safetensors merges…

Dataset · Text to image

UMM-Reflection-SFT-Data

Walt Stephen

The reflection-SFT data of UMM-Reflection (Learning Native Reflection in Unified Models). It trains Each row is one complete trajectory for one request. A T2I trajectory starts with no image; an edit trajectory starts from a given source image. Every controller turn is a block with [CURRENTROUND], [SCORE], [ACTION] (edit or done), [THINKING] and, for an edit, an [EDIT] payload. Each edit is followed by the image it produced. Trajectories contain one to four generated images; they include one-shot successes, planned multi-step progressions, and natural repairs of a flawed image. The rendered [EDIT] line in thinklist can be shortened. The training code uses the full editinstruction from…