Evaluation-Verification Reward for Consistent Multi-Reference Image Editing
Yingmao Miao, Pengfei Zhang, Xiaochen Lv, Meng Yu, Lei Sun, Xiangxiang Chu, Chao Shen, Chenhao Lin
cs.CV
2026-07-31
EVR splits editing scoring into hypothesize-then-verify, grounding each MLLM judgment in visual evidence. Human alignment reaches 0.707 and the tuned Qwen editor wins 67%.
Multi-reference image editing takes a foreground object reference, a background scene reference, and an instruction (put this sofa in this living room), and must fuse the object into the scene while keeping the object consistent and the whole image harmonious. Today's diffusion models still fail at this; the object warps or fights with the scene.
Reinforcement learning would be the obvious lever, and it works well for text-to-image and single-image editing. Extending it to multi-reference editing runs into the reward model: existing reward models are trained on single-image data and cannot judge constraints across several images. Dropping in a multimodal LLM (MLLM) as a zero-shot judge hits a dilemma. Let it write a long chain-of-thought (CoT) and it drifts along its own generated text into hallucination; remove the CoT and it loses the logic needed to cross-reference multiple images.
EVR (Evaluation-Verification Reward) separates reasoning from visual grounding into two stages, so one model never thinks and judges at once.
Scoring runs over five dimensions: reference consistency, scene consistency, harmony, instruction consistency, and visual quality. For each dimension, an MLLM Evaluator generates K=5 independent hypotheses, each a rationale plus a score from 1 to 5. A Verifier then takes each hypothesis, hunts for concrete visual evidence in the image, accepts what it can ground, rejects what it cannot, and marks reliability.
The reasoning behind the split: judging whether a specific claim holds is a far simpler and more visually grounded task than open-ended evaluation. Generating several hypotheses means that even if some carry hallucinations, the valid insights are still captured; verification then filters the hallucinations out, yielding a stable reward. Per dimension, accepted hypotheses are reliability-weighted into a score, and the five scores combine by geometric mean. The geometric mean is deliberate: if any one dimension collapses, it drags the whole reward down, which blocks reward hacking on a single easy axis. The reward plugs into DiffusionNFT with a rank-64 LoRA to fine-tune Qwen-Image-Edit, with no architectural change.
A separate data pipeline builds the training set: GPT generates object descriptions, Z-Image renders them, scene references are generated conditionally on the objects, and instructions cover replacement and insertion, producing 10,000 triplets at N=2 (one object and one scene reference).
After RL fine-tuning with EVR, Qwen-Image-Edit improves sharply on consistency and harmony.
| Dimension | Baseline Qwen-2509 | EVR-tuned |
| Reference consistency | 3.35 | 4.43 |
| Harmony | 2.22 | 2.73 |
| Instruction consistency | 3.27 | 4.01 |
| Total | 0.56 | 0.70 |
Against the newer Qwen-2511 baseline, total rises from 0.59 to 0.70 and reference consistency from 3.66 to 4.37.
The more important question is whether the reward itself is trustworthy. EVR's agreement with human preference reaches 0.707, well above every single-stage alternative: joint CoT 0.473, decoupled direct 0.578, decoupled CoT 0.629, CoT averaging 0.659, and logit-weighted 0.643. Think-then-verify lands closer to humans than any one-shot evaluation.
In the user study, the tuned model wins 67% against Qwen-2511, 57.2% against Flux2, and 58.0% against Gemini, roughly matching or surpassing specialized systems like NanoBanana. Training cost is contained: 4 H20 GPUs, about 150 GPU-hours.
Multi-reference editing is a key step toward image generation becoming genuinely usable, and prompt engineering and architecture tweaks have largely plateaued. This paper shows the bottleneck is the reward signal, not the RL algorithm. Once reasoning and grounding are separated and verification removes hallucinations, an off-the-shelf editor improves substantially with no structural change. The evaluate-then-verify pattern is not specific to image editing; any task that needs fine-grained multimodal judgment but fears hallucination could borrow it.
For anyone doing diffusion RL, it offers a reward construction that fits multi-reference tasks better than CLIP scores or single-image rewards.
The authors flag the ceiling directly: at N of at least 5 references the base model degrades severely, with feature entanglement and incomplete denoising, and EVR can preserve semantic consistency but cannot rescue low-level image fidelity. The method's ceiling is set by the base model. Scoring still depends on subjective MLLM judgments on a 1-to-5 scale; human alignment of 0.707 is higher than baselines but far from perfect. The 10,000 training triplets are entirely synthetic (GPT plus Z-Image), so the distribution may be narrow, and generalization to complex real-world edits is unverified. The authors are from SenseTime and Xi'an Jiaotong University; the base (Qwen-Image-Edit) and the comparison system (NanoBanana) are external, so the conflict is lighter, but matching NanoBanana is a claim made on the authors' own test set.