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
CommerceGen-Studio generates commercial-grade e-commerce main images by jointly modeling product, trademark, and text layout.
How can we automate the generation of high-fidelity, commercially viable e-commerce main images from raw SKU data while maintaining strict trademark and selling-point accuracy?
E-commerce search results require high-quality, SKU-specific imagery, but automated systems often struggle to reconcile conflicting metadata, preserve product identity, and render accurate text within a coherent layout. CommerceGen-Studio solves this by integrating information extraction, layout planning, and multi-condition generation into a single end-to-end framework. It uses a closed-loop agent to distill metadata, a vision-language layout rewriter to plan composition, and a diffusion transformer to render the final image. On the CommerceGen-Bench, the framework matches the performance of large closed-source models and outperforms open-source baselines in instruction following and product integration.
Paper Primer
The system hinges on a three-stage pipeline: the Prompt-Eval-Refine-Loop (PERL) Agent reconciles heterogeneous SKU metadata into a verified selling-point list, the Layout Thinker Optimization (LTO) layer rewrites this list into a spatial blueprint, and the SKU-DiT generator renders the final image. The core move is the Reinforce-Distill-Reinforce (RDR) training strategy: it aligns the model with human preferences, compresses the sampling process for production-grade latency, and performs a final reinforcement round to recover high-frequency details like trademark edges and font strokes.
CommerceGen-Studio achieves commercial-grade performance on hard e-commerce cases.
On the CommerceGen-Bench, the framework scores 9.43–9.45 overall, matching or exceeding closed-source models like GPT-Image-2.5 and significantly outperforming open-source baselines. The system maintains a few-second inference latency while achieving high scores in instruction following and product integration, the two primary metrics for production readiness.
Why is an end-to-end framework necessary instead of just using a general-purpose image editor?
General-purpose models lack the specific constraints required for e-commerce, such as character-level text accuracy, trademark integrity, and the ability to reconcile conflicting metadata from multiple sources like OCR, titles, and backend attributes.
What is the role of the Layout Thinker Optimization (LTO) layer?
LTO acts as an explicit rewriting layer that translates raw selling points into a structured layout caption, ensuring that the generator does not have to guess the placement, hierarchy, or style of text and product elements.
CommerceGen-Studio demonstrates that specialized, end-to-end training—rather than assembling independent components—is the most effective path to achieving the high-fidelity, multi-condition control required for automated e-commerce asset production.
Introduction to CommerceGen-Studio
We introduce CommerceGen-Studio, a three‑stage framework that generates commercial‑grade e‑commerce images from SKU data.
On e‑commerce search result pages, shoppers rely almost entirely on the product image and title to judge relevance. However, merchants rarely provide high‑quality main images at SKU granularity, leading to low click‑through rates and a critical need for automated, commercial‑grade image generation.
CommerceGen-Studio decouples semantic extraction from layout‑aware generation, guaranteeing that the model never hallucinates product details or trademarks while producing commercial‑grade images.
**Figure 1.** Showcase of CommerceGen-Studio. Given a product main image, an SKU image, and structured product information, the system automatically generates commercial-grade images with subject fidelity, trademark consistency, and precise selling-point text rendering.
The shift from template‑based pipelines to a fully generative e‑commerce imagery system unlocks scalable, high‑quality main image production.
Prior Approaches in E-commerce Generation
We compare prior e‑commerce poster and text generation systems across key capabilities.
Prior work on e‑commerce poster generation falls into three strands: systems that synthesize whole posters, methods that render text within images, and reinforcement‑learning approaches that improve diffusion quality.
PosterMaker creates a complete poster from product data but does not ingest heterogeneous SKU metadata nor enforce trademark constraints.
PosterCraft starts from a free‑form text prompt, producing stylized posters without grounding the design in concrete product references.
The CommerceGen-Studio Pipeline
Describes the PERL‑Agent loop, pretraining, and RDR stages that enable faithful e‑commerce image generation.
When the item image advertises a feature that conflicts with the selected SKU, naïve pipelines would copy the wrong fact into the final picture, causing costly hallucinations. The PERL‑Agent loop solves this by iteratively refining a selling‑point list until every dimension of a reference‑free rubric is satisfied.
The agent treats prompt construction as a closed‑loop optimization: a generator drafts candidate selling‑point lists, a judge scores them, an analyzer pinpoints the weakest rubric dimension, and a refiner nudges the prompt just enough to fix that weakness.
OCR from item image yields “Free Shipping”. OCR from SKU image yields “Free Shipping”.
Property list contains {color: “red”, storage: “128 GB”}.
Title reads “Smartphone X”. Category is “Electronics”.
Generator produces candidate list L₀ = [“Free Shipping”, “Smartphone X”, “128 GB”].
Judge scores L₀: SC=1.0, EQ=0.9, CF=0.4 (conflict on storage), RL=1.0, ST=0.9, BV=1.0.
Analyzer flags CF (Category Fitness) as the worst dimension and suggests adding “Storage: 128 GB”.
Refiner appends “Storage: 128 GB” to the prompt, yielding $\pi$₁.
Generator with $\pi$₁ produces L₁ = [“Free Shipping”, “Smartphone X”, “128 GB”, “Storage: 128 GB”]; all rubric scores rise to ≥0.9.
The loop fixes the single failing dimension with a tiny prompt edit, demonstrating why optimizing the minimum score yields rapid convergence.
PERL loop – generate, evaluate, analyze, refine.
**Figure 3.** Prompt-Eval-Refine-Loop Agent (PERL-Agent): the prompt-evaluation-refinement pipeline. Generator produces selling-point text list candidates from the current prompt; Judge scores them against a six-dimensional reference-free rubric (SC: Source Compliance · EQ: Expression Quality · CF: Category Fitness · RL: Red-line Check · ST: Structure Compliance · BV: Business Value); Analyzer attributes low-scoring samples by dominant failure dimension to obtain coverage-tagged patterns; Refiner selects the top-1 pattern and applies a minimal single-point diff to produce the next-round prompt.
Stage I adapts the generic Swift‑Image 6B backbone to the e‑commerce SKU distribution by continual training on white‑background product images and paired captions, filtering samples with aesthetic and saturation models.
Start from the pretrained Swift‑Image 6B model.
Collect triplets (white‑background image, prompt, ground‑truth) from SKU listings.
Filter candidates twice: first with an in‑house aesthetics model, then with a high‑saturation detector.
Sample a balanced subset across leaf categories, scene types, and text length.
Train on ~4 M product‑domain T2I samples, ~2.67 M single‑image I2I pairs, and ~274 K multi‑image I2I pairs.
Result: a backbone that preserves subject fidelity while handling multi‑condition text rendering.
**Figure 2.** Overall framework of CommerceGen-Studio. The training pipeline combines supervised fine-tuning, RDR progressive training, and LTO. The inference pipeline performs selling-point text list extraction, Layout Thinker rewriting, and multi-condition SKU-DiT generation. Dashed lines connect training-time components to their inference-time roles.
Stage II introduces Reinforce‑Distill‑Reinforce (RDR), a three‑phase training regime that first aligns the model with human preferences, then compresses inference cost via distillation, and finally restores fidelity with a second reinforcement pass.
RDR first teaches the model what humans value, then squeezes the inference budget, and finally patches any quality loss caused by the compression.
Layout Thinker Optimization
LTO adds a layout rewriting step and trains it via supervised and reinforcement fine‑tuning.
Feeding the raw TextList straight to the generator forces the model to invent its own layout, which leads to wildly different compositions across otherwise identical samples.
LTO adds a dedicated rewriting layer that turns the raw TextList into a concrete layout caption before generation, stabilizing composition.
Step 1: The VLM reads $L$ and $I_{\text{prod}}$, attends to the shoe region, and predicts the first token “centered”.
Step 2: It conditions on “centered” and predicts “shoe”.
Step 3: Using the same visual context it predicts “bright”.
Step 4: It predicts “lighting”.
Step 5: Finally, conditioned on the previous four tokens and the optional $I_{\text{logo}}$, it predicts “logo top‑right”.
The VLM learns to allocate visual space before any image is rendered, ensuring that subsequent generation respects the prescribed geometry and hierarchy.
How does LTO differ from simply adding more prompt tokens to guide layout?
LTO does not rely on longer prompts; it introduces a separate VLM that explicitly reasons about spatial arrangement and logo placement, producing a structured layout caption that the generator then follows. Plain prompts cannot guarantee consistent geometry because the generator still infers layout implicitly.
Benchmark and Training Data
Dataset, benchmark, and evaluation protocol for e‑commerce image generation.
CommerceGen‑Bench provides a comprehensive testbed for e‑commerce image editing.
It contains 400 real online cases, evenly split into 200 single‑image and 200 multi‑image samples across 6 top‑level categories and 15 sub‑categories.
The training corpus follows a three‑task composition—text‑to‑image, single‑image edit, and multi‑image edit—balanced across product categories, background classes, and selling‑point density. Data cleaning uses an in‑house aesthetic scorer built on SwinV2‑T to filter low‑quality or oversaturated merchant images before sampling. Evaluation resizes all images to ~1024 × 1024 and scores five dimensions (PF, IF, PI, ND, AD) on a 0–10 scale, averaging them for the overall metric.
A curated set of 400 real e‑commerce editing cases that stresses hard, long‑tail scenarios.
The backbone model, a 6‑billion‑parameter DiT, is pretrained on generic image data and then fine‑tuned on SKU‑domain curricula.
**Figure 4.** Task taxonomy of the SKU vertical-domain training corpus. From left to right: three top-level tasks with their corpus share, their sub-tasks with shares, and representative samples.
**Figure 5.** Distribution of CommerceGen-Bench. The benchmark comprises 400 common e-commerce scenarios covering 6 top-level categories and 15 sub-categories, with the percentage share of each category visualized.
Quantitative Performance Evaluation
Our model sets a new overall benchmark on CommerceGen‑Bench.
The core premise remains: generating commercial‑grade e‑commerce images requires a tight balance between creative layout and faithful product rendering. Our experiments evaluate how well each system respects that balance.
Our 6 B SKU‑DiT model achieves the highest overall score on CommerceGen‑Bench, surpassing all open‑ and closed‑source baselines.
Table 1 shows an overall score of 9.37 for our model, compared with 9.22 – 9.71 for the next best closed‑source systems and 8.60 – 8.24 for the best open‑source systems.
Across the board, open‑source baselines lag behind in the IF (Image‑Fidelity) and AD (Attribute‑Diversity) dimensions, while closed‑source systems excel in IF but still fall short of our model’s balanced performance. This demonstrates that our staged full‑parameter + low‑rank training pipeline yields a more uniformly high‑quality output.
Mask2Former provides a unified segmentation head that predicts both class masks and instance masks, enabling precise extraction of product silhouettes even against cluttered backgrounds.
**Table 1.** Quantitative results on CommerceGen-Bench without prompt enhancement (PE). Bold and underline denote the best and second-best results across all open-source and closed-source baselines; Ours is reported separately and excluded from this ranking. Overall is the mean of PF, IF, PI, ND, and AD within each editing track. Rankings are determined from unrounded values; all displayed values are rounded to two decimal places.
Qualitative Analysis
Additional quantitative and qualitative results highlight our model’s superiority on CommerceGen‑Bench.
Our 6B‑parameter model achieves the highest average Single‑Image Editing score on CommerceGen‑Bench without prompt enhancement.
It records 9.66, surpassing the best open‑source baseline (9.55) and all closed‑source baselines.
These numbers confirm that the decoupled semantic‑extraction and layout‑aware generation pipeline scales effectively across both single‑ and multi‑product editing scenarios, while other models struggle with compositional fidelity.
**Figure 6.** Qualitative comparison of CommerceGen-Studio and several models on single-image editing and multi-image editing.
**Figure 7.** Qualitative results of several models without prompt enhancement (PE).
Discussion and Failure Modes
Key limits of performance and data scaling are examined.
Table 3 shows the ablation study on CommerceGen‑Bench. Adding SFT to the VLM base raises the score from 6.51 to 7.83, and incorporating reinforcement (RD, RDR) further improves performance, with LTO VLM leading the table with the top score. The progression demonstrates that each component—SFT, reinforcement, and LTO—contributes incremental gains.
Table 4 reports the data volume for each training stage, ranging from roughly 4 M samples at the initial stage down to 2.5 K for the final fine‑tuning. The steep reduction reflects a curriculum that increasingly focuses on higher‑quality, layout‑rich examples.
In the quantitative comparison (Section 5.2), CommerceGen‑Studio attains overall scores of 9.40 (single‑image) and 9.44 (multi‑image) without prompt enhancement, improving slightly to 9.43 and 9.45 with PE. Open‑source baselines such as Qwen‑Image‑2.1 reach 7.29/6.76 without PE and 8.95/8.90 with PE, still lagging behind in product fidelity, integration, and artifact control.
**Evaluation Prompt (Forensic Rubric Scoring)** You are a conservative, forensic evaluator of commercial product image editing. Compare every visible detail in the generated image against the reference image(s) and the editing instruction. Return ONLY one JSON object. **CALIBRATION RULES** * Score five dimensions independently from 0 to 10; decimals are allowed and higher is better. * Start each dimension at 5.0, then move it only when visible evidence satisfies an anchor. Do not start from 10 and subtract. * 5 means mixed/borderline, not catastrophic. 7 means good but clearly flawed. 8 means strong with limited defects. 9 means excellent with only negligible defects. 10 means essentially flawless after close inspection and must be rare. * Never award 9 or 10 from a favorable overall impression alone. A score above 8 requires naming why there is no material defect in that dimension. * When a relevant small detail or text is visibly unreadable, do not assume it is correct. Do not reward beauty, resolution, or style in dimensions where it does not prove correctness. * Use the lower end of a band when several defects exist and the upper end when only one limited defect exists. When uncertain between two scores, choose the lower score. **DIMENSIONS** * **PF – Product Fidelity:** identity, silhouette, geometry, proportions, materials, colors, fine details, original packaging text, trademarks, and product logos. Layout or camera changes alone are not fidelity errors. * **IF – Instruction Following:** completion and correctness of every explicit requested edit, object, placement, count, relation, style, and newly added or modified text. * **PI – Product Integration:** how coherently the preserved product and optional logo belong in the edited composition, including placement, scale, perspective, lighting, shadows, reflections, edges, occlusion, and scene interaction. * **ND – Naturalness Detection:** perceptual realism and visual plausibility appropriate to the requested photo, advertising, illustration, or graphic-design style. Judge anatomy, physics, texture, lighting, depth, typography placement, and composition, not mere attractiveness. * **AD – Artifacts Detection:** absence of malformed geometry, duplicated or missing parts, garbled pixels or text, halos, seams, smearing, ghosting, unintended objects, broken edges, and generation residue. A higher score means fewer artifacts. **Overall Score:** the arithmetic mean of the five dimension scores, (PF + IF + PI + ND + AD)/5.
Prompts 1–6 exemplify the diverse layout challenges tackled by the system, from preserving intricate clock details to handling complex clothing textures and dynamic background compositions. Each prompt specifies exact visual constraints, text placement, and lighting, testing the model’s ability to maintain product fidelity while adapting the surrounding layout.
LTO VLM delivers the highest ablation score, yet its gains are bounded by the limited data available at later stages and the modest boost from prompt enhancement.
Implementation Details and Ablations
Ablation study quantifies each component’s impact on overall score.
The ablation study isolates the contribution of each training stage by removing it from the full pipeline and measuring the resulting score drop.
Removing SFT reduces the overall score from 9.440 to 6.509, a drop of 2.93 points.
Baseline open‑source layout score 6.509 versus full configuration score 9.440 (Table 3).
Omitting Distillation lowers the score from 9.440 to 8.296, a loss of 1.14 points.
Distillation stage score 8.296 compared with the full score 9.440 (Table 3).
Skipping post‑distillation reinforcement keeps the score at 8.253 instead of 8.271, a 0.02‑point decrease.
After reinforcement the score rises to 8.271 from 8.253 (Table 3).
Removing LTO eliminates the largest late‑stage gain, dropping the score from 9.440 to 8.271, a reduction of 1.17 points.
LTO raises the score from 8.271 to 9.440 (Table 3).