ProductConsistency: Improving Product Identity Preservation in Instruction-Based Image Editing via SFT and RL

2026-06-17 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

The ProductConsistency dataset and associated approach address the challenge of preserving product identity in instruction-based image editing. Current open and closed-source models often fail to maintain fine-grained features, branding, and textual elements in product-centric scenarios, partly due to a lack of specialized datasets. This new work introduces a supervised fine-tuning (SFT) dataset of 87,000 samples and a reinforcement learning (RL) dataset with 869 unique product images. It also proposes the ProductConsistency Benchmark for standardized evaluation. A Cyclic Consistency reward guides RL training by enforcing semantic preservation through caption similarity. Fine-tuning Qwen-Image-Edit-2511 and Flux.1-Kontext-dev with this dataset demonstrated consistent improvements in OCR and Perceptual metrics, with Qwen-Image-Edit-2511 achieving a 5x reduction in character error rate.

Key takeaway

For Machine Learning Engineers developing instruction-based image editing models, you should integrate the ProductConsistency dataset and Cyclic Consistency reward into your training pipelines. This approach directly addresses the critical challenge of preserving product identity and text fidelity, which current models often fail to achieve. By adopting this methodology, you can significantly reduce character error rates and enhance overall visual quality in product-centric applications.

Key insights

A new dataset and RL reward significantly improve product identity preservation and text fidelity in instruction-based image editing.

Principles

• Semantic preservation is key for product identity.
• Data scarcity hinders product-centric image editing.
• Caption similarity can guide RL for visual consistency.

Method

The approach involves supervised fine-tuning (SFT) and reinforcement learning (RL) using a custom dataset. RL training is guided by a Cyclic Consistency reward based on caption similarity.

In practice

• Use ProductConsistency dataset for product editing.
• Apply Cyclic Consistency reward in RL training.
• Benchmark models with ProductConsistency Benchmark.

Topics

• Instruction-Based Image Editing
• Product Identity Preservation
• Supervised Fine-Tuning
• Reinforcement Learning
• Cyclic Consistency Reward
• OCR Fidelity

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Entropy Is Not Enough: Unlocking Effective Reinforcement Learning for Visual Reasoning via Vision-Anchored Token Selection — Visual reasoning RL requires integrating visual sensitivity with token entropy for effective credit assignment.
• BalCapRL: A Balanced Framework for RL-Based MLLM Image Captioning — BalCapRL balances image captioning quality by jointly optimizing correctness, coverage, and linguistic fluency via a novel RL framework.
• Consistent-Inversion: Reverse Consistency Guidance for Structure-Preserving Visual Editing — Consistent-Inversion uses reverse consistency guidance to correct denoising trajectories, preserving structure in text-guided visual editing.
• ProductWebGen: Benchmarking Multimodal Product Webpage Generation — ProductWebGen benchmarks multimodal models, showing editing-based workflows excel in content appeal, while unified models better follow visual instructions.
• AlbumFill: Album-Guided Reasoning and Retrieval for Personalized Image Completion — AlbumFill uses a vision-language model to retrieve identity-consistent references from personal albums for personalized image completion.
• Preserve, Reveal, Expand: Faithful 4D Video Editing with Region-Aware Conditioning — Faithful 4D video editing requires region-aware conditioning to prevent drift and ghosting in diffusion models.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.