EditCaption: Human-Aligned Instruction Synthesis for Image Editing via Supervised Fine-Tuning and Direct Preference Optimization

2026-04-09

AI Summary

This study addresses three prevalent failure modes in current vision-language models when generating image-editing instructions: directional inconsistency, ambiguous perspective, and insufficient attribute description—collectively causing nearly 48% of automatically generated instructions to contain critical errors, rendering them unsuitable for training. To tackle these issues, the work proposes a two-stage post-training pipeline: first, constructing a high-quality supervised fine-tuning dataset of 100,000 examples through GLM-based auto-annotation, EditScore filtering, and human correction; second, collecting 10,000 human preference pairs to enhance alignment via Direct Preference Optimization (DPO). Experiments based on Qwen3-VL demonstrate substantial improvements, achieving a score of 4.712 on Eval-400, reducing critical error rates from 47.75% to 23%, and increasing correctness to 66%, significantly outperforming open-source baselines.

Abstract

High-quality training triplets (source-target image pairs with precise editing instructions) are a critical bottleneck for scaling instruction-guided image editing models. Vision-language models (VLMs) are widely used for automated instruction synthesis, but we identify three systematic failure modes in image-pair settings: orientation inconsistency (e.g., left/right confusion), viewpoint ambiguity, and insufficient fine-grained attribute description. Human evaluation shows that over 47% of instructions from strong baseline VLMs contain critical errors unusable for downstream training. We propose EditCaption, a scalable two-stage post-training pipeline for VLM-based instruction synthesis. Stage 1 builds a 100K supervised fine-tuning (SFT) dataset by combining GLM automatic annotation, EditScore-based filtering, and human refinement for spatial, directional, and attribute-level accuracy. Stage 2 collects 10K human preference pairs targeting the three failure modes and applies direct preference optimization (DPO) for alignment beyond SFT alone. On Eval-400, ByteMorph-Bench, and HQ-Edit, fine-tuned Qwen3-VL models outperform open-source baselines; the 235B model reaches 4.712 on Eval-400 (vs. Gemini-3-Pro 4.706, GPT-4.1 4.220, Kimi-K2.5 4.111) and 4.588 on ByteMorph-Bench (vs. Gemini-3-Pro 4.522, GPT-4.1 3.412). Human evaluation shows critical errors falling from 47.75% to 23% and correctness rising from 41.75% to 66%. The work offers a practical path to scalable, human-aligned instruction synthesis for image editing data.

Problem

• instruction synthesis
• image editing
• vision-language models
• training data quality
• human alignment

Innovation

• instruction synthesis
• image editing
• direct preference optimization
• supervised fine-tuning
• vision-language models