Research map / Image generation / Text-to-image diffusion

SpatialFlow-GRPO: Where Spatial Credit Drives Image Editing

NeurIPS 2026 · Image generation › Text-to-image diffusion · approach: Preference optimization for diffusion · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SpatialFlow-GRPO, an online RL method for image editing that assigns region-level rewards (from a new region-aware reward model SFReward) to semantic regions and aligns region advantages and policy ratios with latent positions, improving over Flow-GRPO on GEdit, ImgEdit and a new MultiEditBench.

Problem

Flow-GRPO-style online RL for image editing uses a single whole-image reward, implying spatial uniformity of credit, which hampers fine-grained and multi-region editing optimization.

Approach

Replace the sample-level reward with semantic-region rewards from a region-aware evaluator, normalize advantages within (instruction, region label) groups, mix with a global advantage under a warmup schedule, and aggregate policy ratios and losses per latent region with power-weighted region aggregation.

Key results

• On OmniGen2, SpatialFlow-GRPO reaches GEdit overall 7.29 (+0.87) vs Flow-GRPO 6.99 (+0.57) and ImgEdit 3.74 (+0.30) vs 3.66.
• On FLUX.2-klein-4B, SpatialFlow-GRPO gives GEdit 6.80 (+0.22) vs Flow-GRPO 6.73 (+0.15) and ImgEdit 3.53 vs 3.50.
• On MultiEditBench, OmniGen2 + SpatialFlow-GRPO averages 8.11 (+1.70) versus 7.52 for Flow-GRPO; FLUX.2-klein-4B + SpatialFlow-GRPO averages 8.04 vs 7.97 for Flow-GRPO.

Surprising findings

• Adding region advantages alone (without global mixing) does not stably outperform Flow-GRPO, as reported in the progressive ablation.
• On MultiEditBench, the OmniGen2 + SpatialFlow-GRPO variant (8.11) scores below the Qwen-Image-Edit-2509 reference (8.33) in the average, and the 2-edit score is slightly lower than the FLUX.2-klein-4B variant at some subsets.

Limitations

• Relies on region annotations and reliability of the reward model.
• Human evaluation at larger scale is left for future work.
• Pure region advantages can amplify local noise without global constraints.

Experimental setup

Datasets: EditScore-RL-Data SFReward-14K EditReward-Bench MMRB2 GEdit-Bench ImgEdit-Bench

Benchmarks: GEdit-Bench ImgEdit-Bench MultiEditBench EditReward-Bench MMRB2

Metrics: GEdit overall score ImgEdit overall score MEB score (sqrt of SC x PQ) prompt following, consistency, overall accuracy

Baselines: Flow-GRPO Qwen-Image-Edit-2509 Step1X-Edit FLUX.1-Kontext BAGEL-7B-MoT EditScore-7B

Builds on: Flow-GRPO DanceGRPO VIEScore

Closest papers by meaning

• SpatialReward: Bridging the Perception Gap in Online RL for Image Editing via Explicit Spatial Reasoning (ICML 2026 · similarity 0.90)
• CoCoEdit: Content-Consistent Image Editing via Region Regularized Reinforcement Learning (ICML 2026 · similarity 0.89)
• Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.86)
• Edit-R2: Context-Aware Reinforcement Learning for Multi-Turn Image Editing (NeurIPS 2026 · similarity 0.86)
• Multi-Reward as Condition for Instruction-based Image Editing (ICLR 2025 · similarity 0.85)
• Policy Optimized Text-to-Image Pipeline Design (NeurIPS 2025 · similarity 0.85)