PaperlayerView exact revision on arXiv

Source-linked AI summary

CommerceGen-Studio: An End-to-End Framework for Commercial-Grade Main-Image Generation

Xiaoli Xu, Zhengtao Wu, Zhengze Xu, Tao Liu, Yongchao Du, Zuan Gao, Xingjian Wang, Hao Yan, Mengting Chen, Wei Li, Bo Zheng

arXiv:2610.20603v1cs.CV

TL;DR

SKU-level e-commerce images must reconcile heterogeneous product information while preserving accurate text, product identity, and trademarks. CommerceGen-Studio combines metadata extraction, layout rewriting, and multi-condition generation. It achieves the highest Overall scores among evaluated open-source baselines on both editing tracks, with and without prompt enhancement.

  • Problem

    SKU image generation must reconcile heterogeneous metadata and preserve subject, trademark, and text consistency, requirements existing methods address only partially.

    Source: §1 · p. 1

  • Method

    CommerceGen-Studio extracts selling points from SKU metadata, rewrites them as layouts, and generates images using product and trademark references with SKU-DiT.

    Source: §1 · p. 3, §3.1 · p. 5, §3.2 · p. 6

  • Results

    CommerceGen-Studio achieves the highest Overall scores among evaluated open-source baselines on both editing tracks, with and without prompt enhancement.

    Source: §6 · p. 15

  • Takeaways & Limitations

    The framework targets directly deployable SKU main images whose text, product attributes, trademarks, and layout meet specified consistency and quality requirements.

    Source: §3.1 · p. 6, §6 · p. 15

    For Ours, cross-setting Instruction Following changes reflect both generation quality and prompt difficulty, so they are not a pure capability comparison.

    Source: §5.2 · p. 15

Abstract

from arXiv · show

We present CommerceGen-Studio, an end-to-end framework for generating commercial-grade main images for e-commerce Stock Keeping Units (SKUs). Our framework is organized along a three-stage pipeline of information extraction, semantic rewriting, and multi-condition generation: Prompt-Eval-Refine-Loop Agent (PERL-Agent) distills heterogeneous SKU metadata into concise selling-point text lists through an iterative closed loop of VLM and LLM roles scored by a reference-free rubric. Layout Thinker Optimization (LTO) aligns the rewriter with both caption and rendered-image quality via alternating domain-specific policy optimization. SKU-DiT, trained under Reinforce-Distill-Reinforce (RDR) progressive training, conditions jointly on the white-background product image, the trademark crop, and the structured layout to hold subject fidelity, trademark integrity, and character-level text accuracy together. The framework runs in production at a scale of millions of main images per day with a latency of a few seconds per image, which we believe makes it among the first systems of this kind to operate at commercial scale. We also release CommerceGen-Bench, a benchmark of hard cases collected from live traffic across all product categories. On it, our framework achieves the highest overall scores among the evaluated open-source baselines on both editing tracks, with and without prompt enhancement. It also matches much larger closed-source models, and beats several of them at Instruction Following and Product Integration - the two properties that determine whether an image can be published.

1 Introduction

CommerceGen-Studio addresses the need for commercial-quality main images at SKU granularity, where existing merchant imagery and general-purpose image models do not meet key requirements. It combines metadata extraction, semantic rewriting, and multi-condition generation to produce SKU images.

1 Introduction

  • SKU-level images matter because search results increasingly resolve queries to specific variants, while merchants rarely maintain those display assets.Plain white-background SKU photographs used directly in search results measurably lower click-through rate.Source: §1 · p. 1
  • General-purpose image models fall short on reconciling heterogeneous SKU information, consistency, long-tail robustness, and joint product-trademark-text layout control.Existing cascaded systems can accumulate errors across stages, while end-to-end poster models may trade text accuracy against subject fidelity.Source: §1 · p. 1
  • The target image combines subject fidelity, trademark consistency, and precise selling-point text rendering from product and SKU references.The framework overview also distinguishes training and inference pipelines for these capabilities.Source: Figure 1 · p. 2, Figure 2 · p. 3, §1 · p. 3
  • CommerceGen-Studio organizes information extraction, semantic rewriting, and multi-condition generation into an end-to-end framework for SKU main images.Its contributions include PERL-Agent, LTO, SKU-DiT with RDR training, and CommerceGen-Bench.Source: §1 · p. 3

2 Related Work

Prior e-commerce poster and image-editing systems address portions of SKU image creation, but often lack heterogeneous metadata reconciliation, multi-reference conditioning, or coordinated layout control. Related work also motivates explicit text-layout conditioning and specialized optimization and distillation for diffusion models.

2.1 E-commerce Poster Generation

  • Poster and editing systems support composition, typography, or reference-guided editing, but generally omit some combination of raw SKU metadata reconciliation, trademark references, and adaptive multi-reference layout planning.AutoPP uses prepared candidate copy; PosterVerse is prompt-driven, while PosterText and TransAnyText focus on editing or localization.Source: §2.1 · pp. 3–4

2.2 Visual Text Rendering

  • Text-rendering methods improve glyph control or layout flexibility, yet the cited systems do not jointly model text layout with product subject and trademark conditions.The paper positions explicit text layout as a condition path in a unified diffusion Transformer.Source: §2.2 · p. 4

2.3 Reinforcement Learning and Distillation for Diffusion Models

  • Diffusion reinforcement methods face flow-matching or multi-task challenges; the paper adopts DiffusionNFT and proposes dynamic backward-simulation distillation for few-step quality, structural integrity, and trademark generation.DiffusionNFT is described as compatible with flow matching; the proposed distillation increases stopping probability near the clean-image endpoint and adds discriminator-based adversarial training.Source: §2.3 · pp. 4–5, Figure 3 · p. 5

3 Method

CommerceGen-Studio turns heterogeneous SKU metadata into selling points, rewrites them as executable layouts, and renders images conditioned on the product and trademark references. Its training pipeline combines domain adaptation, progressive generator optimization, and layout-rewriter training.

3.1 Overview

  • The inference pipeline extracts a TextList from SKU metadata, rewrites it as a layout caption, then conditions SKU-DiT on that caption, the white-background product image, and trademark crop.The intended output preserves metadata-supported selling points and their exact text, product attributes, trademark integrity, and typographic coherence.Source: §3.1 · pp. 5–6

3.2 PERL-Agent for TextList Extraction

  • PERL-Agent iteratively refines a category-adaptive prompt using VLM-generated candidates and LLM evaluation roles, optimizing the weakest dimension of a six-dimensional reference-free rubric.The Analyzer identifies a dominant failure and the Refiner makes a minimal, single-point prompt modification; intermediate artifacts support comparison and rollback.Source: §3.2 · p. 6

3.3 Stage I: Pretraining & Continual Training

  • Continual training adapts the Swift-Image 6B backbone using about 4M product-domain T2I samples, 2.67M single-image editing pairs, and 274K multi-image editing pairs.The data are filtered and balanced across leaf categories, background scenes, and in-image text lengths.Source: §3.3 · p. 6

3.4 Stage II: Progressive Reinforce-Distill-Reinforce Training

  • RDR aligns the generator with human preferences, distills it for few-step inference, then reinforces the student to recover fidelity and text precision under the few-step constraint.Distillation combines general and trademark-domain editing data at a 3:1 ratio and brings single-image latency to the 5-second level; post-distillation reinforcement uses roughly 2.5K grouped trajectories.Source: §3.4 · p. 7

3.5 Stage III: Layout Thinker Optimization

  • LTO trains a VLM to rewrite TextLists into layout captions, then optimizes it against caption and rendered-image rewards while activating only one reward domain per step.SFT uses balanced sampling; during RFT, SKU-DiT remains frozen as the renderer and only the VLM policy is updated.Source: §3.5 · pp. 8–9

4 Dataset and CommerceGen-Bench

The section describes the SKU-domain training data, the hard-case benchmark, and the evaluation protocol used to assess CommerceGen-Studio.

4.1 Training Model

  • The generation backbone is Swift-Image 6B, a unified multi-image editing DiT initialized from its reported 500K-step pretraining and 200K-step continual-training checkpoint.Training resolution increases from 256 to 1024 pixels through a five-stage curriculum before SKU-domain adaptation.Source: §4.1 · p. 9

4.2 Training Data

  • Training data include complex-layout and subject-preservation editing samples, with balanced sampling across product categories, 50 background classes, and selling-point counts.Subject-preservation samples require product identity, structure, color, texture, and quantity to remain unchanged during editing.Source: §4.2 · p. 9

4.3 CommerceGen-Bench

  • CommerceGen-Bench contains 400 live-search cases, evenly split between single- and multi-image editing, spanning 6 top-level categories and 15 subcategories.Hard cases include thin or transparent subjects, small trademarks near the resolution limit, or long selling-point text prone to character errors.Source: §4.3 · p. 10, Figure 5 · p. 10

4.4 Evaluation Protocol

  • A multimodal judge assesses generated images against reference products and instructions on five dimensions scored from 0 to 10.The dimensions are Product Fidelity, Instruction Following, Product Integration, Naturalness Detection, and Artifacts Detection; Overall is their arithmetic mean.Source: §4.4 · pp. 10–11

5 Experiments

Experiments compare CommerceGen-Studio with open- and closed-source models, examine qualitative outputs, and ablate training stages. Results report strong benchmark scores and gains associated with the training components.

5.1 Training Configuration

  • The configuration alternates full-parameter training for capability changes with low-rank reinforcement updates to limit drift from prior grounding.The authors identify unconstrained reward optimization as a risk to subject and text fidelity.Source: §5.1 · p. 11

5.2 Quantitative Comparison

  • CommerceGen-Studio scores 9.40 and 9.44 Overall without prompt enhancement, and 9.43 and 9.45 with it, on the single- and multi-image tracks, respectively.Among open-source baselines, Qwen-Image-2.1 is strongest overall in both settings, but remains below CommerceGen-Studio in product fidelity, integration, and artifact control.Source: Table 1 · p. 12, Table 2 · p. 12, §5.2 · p. 14
  • Prompt enhancement substantially improves instruction following for most baselines, while domain-specific training remains critical for faithful product-and-trademark composition.For CommerceGen-Studio, the enhanced prompt adds explicit placement, hierarchy, and style requirements, increasing the number of verifiable constraints.Source: §5.2 · p. 15

5.3 Qualitative Examples

  • Figures 6 and 7 show more consistent preservation of product appearance, logos, and selling-point text, while illustrating identity drift, corrupted text, and unstable multi-reference composition in other models.Figure 6 covers single- and multi-image editing; Figure 7 presents results without prompt enhancement.Source: Figure 6 · p. 13, Figure 7 · p. 14, §5.3 · p. 15

5.4 Ablation Study

  • LTO raises the ablation score from 8.271 to 9.440, the largest late-stage gain; distillation enables 8-step sampling while the score decreases from 8.296 to 8.253.Post-distillation reinforcement partially recovers the score to 8.271.Source: Table 3 · p. 12, §5.4 · p. 15

6 Conclusion and Future Work

CommerceGen-Studio combines extraction, layout rewriting, and generation to produce SKU-granular main images. On CommerceGen-Bench, it leads evaluated open-source baselines on both editing tracks, while the authors identify limitations in reward signals and personalization.

6 Conclusion and Future Work

  • CommerceGen-Studio achieves the highest Overall scores among evaluated open-source baselines on both editing tracks, with and without PE, and outperforms several closed-source systems on selected dimensions.It remains competitive with leading closed-source systems.Source: §6 · p. 15
  • PERL-Agent extracts selling points from heterogeneous metadata, Layout Thinker Optimization creates executable layouts, and SKU-DiT renders them in a few seconds without surrendering fidelity.The framework produces commercial-grade main images at SKU granularity.Source: §6 · p. 15
  • The authors identify rubric-defined rewards and a single quality target per SKU as limitations, proposing real-user behavior and audience- or style-conditioned assets as future directions.Real-user behavior could capture aesthetic preferences that resist explicit specification.Source: §6 · p. 15

A.1 Evaluation Prompt

Figure 8 presents an abridged forensic evaluation prompt scored with a conservative five-dimensional rubric. The overall score is the arithmetic mean of its dimension scores.

A.1 Evaluation Prompt

  • Each dimension has anchor descriptions for the 10, 9, 7–8, 5–6, 3–4, 1–2, and 0 score bands.The anchors are written in the same conservative style as the calibration rules.Source: §A.1 · p. 15
  • The abridged prompt omits per-band anchor descriptions while retaining instructions that govern scoring behavior.The anchor tables are omitted for space.Source: §A.1 · p. 15, Figure 8 · p. 16
  • The forensic rubric scores five dimensions—PF, IF, PI, ND, and AD—from 0–10, and defines the overall score as their arithmetic mean.Figure 8 reproduces the judge model's evaluation prompt in abridged form.Source: Figure 8 · p. 16

A.2 PERL-Bench-6

PERL-Agent TextList quality is evaluated with a ground-truth-free, six-dimensional rubric. Each dimension is scored from 0 to 1, and a low aggregate score triggers analysis and rollback.

A.2 PERL-Bench-6

  • Source Compliance, Expression Quality, and Category Fitness assess metadata consistency, language quality, and category-specific fit, respectively.Each dimension is scored from 0 to 1.Source: §A.2 · p. 16
  • Red-line Check detects rule violations or sensitive material, Structure Compliance measures text-list regularity, and Business Value measures selling-point conversion potential.These are the rubric's remaining three dimensions.Source: §A.2 · p. 16
  • An aggregate score of 5 or lower out of 6 is classified as a bad case, triggering Analyzer attribution and Refiner rollback.A rule module applies reverse penalty scoring to specific patterns.Source: §A.2 · p. 16

Related papers

Loading 2610.20603v1…