Research map / Vision-language and multimodal / Multimodal LLMs

UnicEdit-10M: A Dataset and Benchmark Breaking the Scale-Quality Barrier via Unified Verification for Reasoning-Enriched Edits

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 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.

Builds UnicEdit-10M, a 10M-triplet image-editing dataset curated by an end-to-end editing model plus a unified post-verification stage run by a 7B expert (Qwen-Verify), and UnicBench, a 22-task benchmark with NC and RA metrics.

Problem

Open-source image editing lags closed-source models because large high-quality training data and benchmarks that diagnose complex (spatial, reasoning) edits are scarce; existing automated pipelines suffer error propagation and noisy, misaligned triplets.

Approach

Replace multi-tool pipelines with end-to-end editing models (FLUX.1-Kontext, Qwen-Image-Edit), then apply a unified post-verification stage in which a 7B dual-task expert (Qwen-Verify, SFT + D2PO preference optimization) filters failed edits and recaptions instructions; evaluate with a benchmark adding Non-edit Consistency and Reasoning Accuracy metrics.

Key results

• Qwen-Verify (7B) reaches 6.32 Normal, 9.80 No Edit, 6.22 Hallucination accuracy, beating Qwen2.5-VL-72B (5.25, 9.60, 6.12) on its expert test set.
• On UnicBench, GPT-Image-1 has the top overall score (8.6830 EN) and Qwen-Image-Edit leads open-source models (7.7273 EN overall), with nearly all models showing a sharp drop on Reasoning Accuracy.
• UnicEdit-10M achieves face consistency 0.8911 versus 0.3025 for GPT-Image-Edit-1.5M, and the highest PQ and overall score among compared datasets (Overall 7.76).

Surprising findings

• Nearly all evaluated models, including strong closed-source ones, show a significant drop on Reasoning Accuracy relative to basic instruction following.
• SSIM-based no-edit filtering is unreliable: it gives high scores to visible small edits and low scores to unchanged pairs due to pixel noise.

Limitations

• Verification relies on GPT-4.1 as judge for benchmark and expert-model evaluation, so scores depend on a proprietary model.
• Expert model training data partly generated and corrected with GPT-4o and human annotators; scale of human involvement is limited to a few hundred test samples.
• Dataset generation depends on internal image library whose release status is unclear.

Experimental setup

Datasets: UnicEdit-10M ImgEdit SEED-Data-Edit GPT-Image-Edit-1.5M

Benchmarks: UnicBench VIEScore comparison on 1K sampled triplets

Metrics: Instruction Following Non-edit Consistency Visual Quality Reasoning Accuracy Alignment Accuracy VIEScore Aesthetic score facial consistency

Baselines: Qwen2.5-VL-72B Qwen2.5-VL-7B SSIM filtering VIEScore GEdit-Bench metric GPT-Image-Edit-1.5M ImgEdit SEED-Data-Edit

Builds on: Qwen2.5-VL FLUX.1-Kontext Qwen-Image-Edit DPO VIEScore

Closest papers by meaning

• KRIS-Bench: Benchmarking Next-Level Intelligent Image Editing Models (NeurIPS 2025 · similarity 0.91)
• ReasonEdit: Towards Reasoning-Enhanced Image Editing Models (CVPR 2026 · similarity 0.90)
• Learning Action and Reasoning-Centric Image Editing from Videos and Simulation (NeurIPS 2024 · similarity 0.90)
• EdiVal-Agent: An Object-Centric Framework for Automated, Fine-Grained Evaluation of Multi-Turn Editing (ICLR 2026 · similarity 0.89)
• Envisioning Beyond the Pixels: Benchmarking Reasoning-Informed Visual Editing (NeurIPS 2025 · similarity 0.89)
• Reasoning to Edit: Hypothetical Instruction-Based Image Editing with Visual Reasoning (ICML 2026 · similarity 0.89)