Current breakthroughs in instruction-based picture enhancing have captured vital consideration, as fashions at the moment are able to dealing with real-world enhancing calls for with the practicality required by on a regular basis customers. Nonetheless, enhancing fashions skilled primarily for single-turn edits usually break down in multi-turn enhancing—the pure interactive setting the place a person iteratively refines a picture based mostly on the mannequin’s personal earlier outputs. This failure stems from the all-or-nothing requirement, the place a single failed flip compromises your entire sequence, and error propagation, the place publicity bias results in compounding enhancing errors. To handle these challenges, we introduce MT-EditFlow, a flow-matching reinforcement studying framework designed to optimize reward indicators for sequential picture enhancing. MT-EditFlow integrates a multi-turn perspective with a multi-reward formulation to supply a unified construction relevant to each GRPO and NFT-based reinforcement studying strategies. We systematically analyze and optimize the reward sign by investigating efficient scoring methods for turn-level aggregation, VLM reasoning modes to commerce off reward bias and variance, and benefit fusion ranges to forestall reward hacking. Our findings reveal that broadcasting the aggregated benefit throughout your entire enhancing trajectory successfully bridges the hole between native planning and world multi-turn process success. Intensive experiments reveal that MT-EditFlow considerably improves efficiency throughout various base fashions. Notably, it boosts FLUX.1-Kontext-dev by 6.85 factors in turn-3 total efficiency, surpassing state-of-the-art open-source fashions similar to Qwen-Picture-Edit. By sustaining excessive marginal success charges and decreasing publicity bias, MT-EditFlow supplies a basis for extra dependable and pure human-AI collaboration in visible content material creation.
- †College of California, Los Angeles
- ‡ College of Texas at Austin
- § Lambda, Inc
- * Equal contribution
