💻 computer science

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

This paper introduces the ProductConsistency dataset and a Cyclic Consistency reward mechanism to enhance product identity preservation in instruction-based image editing, demonstrating that supervised fine-tuning and reinforcement learning significantly improve text fidelity and visual quality in models like Qwen-Image-Edit-2511 and Flux.1-Kontext-dev.

Original authors: Mukund Khanna, Raj Singh Yadav, Kunal Singh

Published 2026-06-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mukund Khanna, Raj Singh Yadav, Kunal Singh

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a favorite coffee mug with your name and a specific logo printed perfectly on it. You want to use an AI tool to take a photo of that mug and place it on a sunny kitchen counter for a social media post.

Ideally, the AI should move the mug, change the background, and maybe add some sunlight, but keep your name and the logo exactly as they are.

The Problem: The "Clumsy AI" Artist
The paper explains that current AI image editors are like very talented but clumsy artists. They are great at following instructions like "put this on a shelf" or "change the lighting." However, when it comes to keeping specific details of a product (like a brand logo, a specific font, or the exact shape of a bottle) intact, they often mess up.

Instead of keeping your name on the mug, the AI might:

Turn your name into gibberish.
Change the logo's color.
Distort the shape of the bottle.
Invent new, fake text that wasn't there before.

For businesses (like advertisers or online stores), this is a disaster. You can't sell a product if the AI accidentally changes the brand name or makes the packaging look weird.

The Solution: A Specialized Training Camp
The authors created a new system called ProductConsistency to fix this. Think of it as a specialized training camp for AI artists where the only rule is: "Do not change the product's identity."

Here is how they did it, using simple analogies:

Building a Massive Practice Library (The Dataset):
The AI needs to practice on thousands of examples to learn. The authors built a huge library of 87,000 synthetic product images. They used a computer program to generate fake products (like a soda can or a cereal box) with perfect, readable text and unique logos.
- The Filter: Since AI sometimes writes gibberish when generating text, they used a "spell-checker" (OCR) to throw away any image where the text wasn't perfect. Only the images with perfect text made it into the training library.
Two-Step Training (SFT and RL):
They trained the AI in two phases:
- Phase 1: Supervised Fine-Tuning (SFT) - "The Homework": The AI looked at the perfect images and learned how to copy them while changing the background. It's like a student memorizing the correct answers to a test.
- Phase 2: Reinforcement Learning (RL) - "The Coach's Whistle": This is where the real magic happened. They introduced a new "reward system." Imagine a coach who doesn't just look at the final picture, but also checks if the essence of the product is still there.
  - They invented a "Cyclic Consistency Reward." Here's how it works: After the AI edits the image, another AI (a "caption writer") describes the new image. The system then compares this new description to the original product description. If the descriptions match closely (meaning the product still looks and sounds like the same product), the AI gets a high score. If the AI changed the logo or the text, the descriptions won't match, and the AI gets a low score.
The Result: A Master Editor
They tested this new, trained AI against other top models. The results were impressive:
- Text Accuracy: The AI made 5 times fewer spelling errors on product text compared to the standard models.
- Brand Safety: The logos and shapes stayed true to the original.
- Overall Quality: Even human-like AI judges (Large Language Models) rated the new models higher for keeping the product looking "real" and consistent.

In Summary
The paper argues that to make AI useful for real-world product marketing, we can't just rely on general image editing tools. We need to specifically train them on data that cares about brand identity and text accuracy. By creating a massive library of perfect product images and teaching the AI a "consistency check" (the Cyclic Consistency reward), they successfully taught the AI to edit photos without ruining the product's identity.

What the paper does NOT claim:

It does not claim this will be used for medical imaging or clinical diagnosis.
It does not claim this will replace human graphic designers entirely, only that it improves the specific task of editing product photos.
It does not claim this works for every type of image, only for instruction-based editing of products where brand consistency is key.

Technical Summary: ProductConsistency

Problem Statement

While recent advances in instruction-based image editing have enabled complex visual modifications from natural language, current models (both open and closed source) struggle significantly in product-centric scenarios. In domains such as advertising, e-commerce, and marketing, edited images require strict visual correctness. Existing models frequently fail to preserve fine-grained object identity, often distorting product geometry, altering brand logos, or hallucinating and misspelling text on product surfaces.

A primary driver of this failure is the lack of dedicated datasets and benchmarks that explicitly target brand consistency and text fidelity during editing. Existing benchmarks often ignore rendered text or focus on document-style layouts rather than photographed objects. Furthermore, current reinforcement learning (RL) reward models are designed for general aesthetic quality or instruction following, failing to capture the specific constraints of product identity preservation.

Methodology

The authors propose a comprehensive framework comprising a new dataset, a synthetic data generation pipeline, and a specialized training strategy involving Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL).

1. ProductConsistency Dataset and Generation Pipeline

The core of the work is a fully automated pipeline for generating high-quality synthetic product images with verifiable text and unique brand identities.

Synthetic Generation: The pipeline defines 131 unique product items across 8 categories (e.g., food, beverages, cosmetics). An MLLM (GPT-o3-mini) generates diverse product descriptions and brand signatures (name, color palette, font style).
Image Synthesis & Filtering: Images are generated using GPT-Image-1 High. To ensure text fidelity, a rigorous OCR-based filtering stage is applied. An image is retained only if the text detected by an OCR model ( $T_{ocr}$ ) exactly matches the ground-truth text specified in the prompt ( $T_{gt}$ ). This process yields a high-precision set of 2,002 synthetic products with legible text.
Data Splitting: The filtered data is split into:
- SFT Dataset: 87,242 samples used for supervised fine-tuning.
- RL Dataset: 869 unique product images (derived from 1,133 unique products) selected based on high variance in baseline model performance, representing "failure-prone" cases where optimization is most beneficial.
- Benchmark: A human-verified test set of 174 product images (870 total samples with 5 prompts each) across 8 categories.

2. Training Framework

The authors fine-tune two open-source models: Qwen-Image-Edit-2511 and Flux.1-Kontext-dev.

Supervised Fine-Tuning (SFT): Models are trained on the 87k SFT dataset using Low-Rank Adaptation (LoRA). Target images for SFT are generated using Qwen-Image-Lightning (chosen for speed and text rendering) and filtered via OCR to ensure only correct-text outputs are used as ground truth.
Reinforcement Learning (RL): Training continues from SFT checkpoints using the FlowGRPO algorithm.
- Cyclic Consistency Reward: To guide RL without explicit product-consistency reward models, the authors propose a proxy reward. They generate a caption ( $c_{gen}$ ) from the edited image using an MLLM (Qwen3-VL) and compute the cosine similarity between its embedding and the embedding of the original product description ( $c_{gt}$ ). This enforces semantic preservation of product identity.
- OCR Reward: This is combined with a Character Error Rate (CER) based reward to explicitly penalize text hallucinations and spelling errors.
- Aggregation: The GDPO algorithm is used to aggregate the cyclic consistency and OCR rewards with equal weight.

Key Contributions

ProductConsistency Dataset & Pipeline: A fully automated pipeline for generating synthetic product images with unique brand identities and verifiable rendered text, enabling scalable supervision for product-centric editing.
ProductConsistency Benchmark: A rigorous evaluation suite consisting of 174 product images across 8 categories, paired with 5 editing prompts each (870 samples), designed to evaluate product identity preservation, text fidelity, and visual consistency.
Product-Aware Training with Cyclic Consistency Rewards: A novel RL framework that aligns generated image captions with original product descriptions and incorporates OCR-based rewards. This approach explicitly targets instruction adherence, semantic product identity, and text fidelity.

Results

The authors evaluated their fine-tuned models against strong baselines (including HiDream-E1-1, OmniGen2, NanoBanana, and GPT-Image-1) on the ProductConsistency Benchmark using OCR metrics, segmentation-based perceptual metrics (Seg CLIP-I, Seg DINO-I), and MLLM-based evaluations (GPT-5.1).

Quantitative Improvements: Models fine-tuned with the ProductConsistency dataset showed consistent improvements across all metrics.
- Qwen-Image-Edit-2511: SFT reduced the Character Error Rate (CER) from 1.0682 to 0.7803. The addition of RL with Cyclic Consistency rewards further reduced CER to 0.2080, representing a 5× reduction in character error rate compared to the base model. Seg CLIP-I improved from 0.8728 to 0.9245.
- Flux.1-Kontext-dev: Starting from a strong baseline, the model achieved the lowest overall CER of 0.1204 and the best DINO-I score of 0.8115 after SFT and RL training.
Qualitative & MLLM Evaluation: GPT-based evaluations confirmed that the improvements in automated metrics translated to perceptible gains in product consistency, text fidelity (including font and color), and overall aesthetics.
Ablation Study: The study compared the Cyclic Consistency reward against EditReward and Segmented Visual Consistency rewards. While Segmented Visual Consistency achieved high automated scores, it suffered from "reward hacking," where models reproduced the input image rather than performing edits, leading to poor aesthetics and instruction following. The Cyclic Consistency reward provided a more robust training signal, outperforming other configurations in both automated and GPT-based evaluations.

Significance

The paper claims that its work addresses a critical gap in the field: the inability of current instruction-based editing models to maintain fine-grained product identity and text fidelity. By introducing a dedicated dataset, benchmark, and a training framework centered on cyclic consistency and OCR rewards, the authors demonstrate that explicitly modeling product consistency is essential for real-world applications in advertising and e-commerce. The results indicate that fine-tuning strong open-source editors with this specific framework yields substantial gains, making them competitive with or superior to closed-source models in product-centric editing tasks.