KwaiMind: Kuaishou's E-Commerce Image Editor Learns User Preferences, Boosts CTR 2.44%

KwaiMind: Unified E-Commerce Image Editing Foundation Model

Kuaishou's technical team introduced KwaiMind, a foundation model for real-world e-commerce content production that merges general image editing with specialized e-commerce capabilities. The model handles addition, removal, replacement, composition adjustment, and local defect repair within a single instruction-based editing framework, supporting tasks such as virtual try-on, product detail showcase, accessory placement, background replacement, and pose adjustment.

Multi-Agent Data Engine

To handle massive, diverse e-commerce material while ensuring quality, KwaiMind employs a multi-agent data engine that integrates quality filtering, sample repair, targeted supplementation, and instruction annotation into a unified pipeline with automatic task routing and sample traceability. Four agents collaborate:

Filter Agent: Pre-screens input images and re-screens editing results.

Generation Agent: Repairs recoverable samples per diagnosis and supplements underrepresented task data.

Caption Agent: Generates editing instructions at multiple granularities.

Coordinator Agent: Maintains sample state and distribution, orchestrates task scheduling, process handoff, and retries.

A Human-in-the-Loop mechanism routes low-confidence, high-disagreement samples to human reviewers, with results fed back into the pipeline. From ~26.2 million candidate pairs, the engine constructs and maintains ~1.8 million high-quality training pairs for continued fine-tuning, and further selects ~234,900 pairs for SFT (115k general editing, 119.9k e-commerce). General data solidifies base editing ability; e-commerce data strengthens garment detail, product display, and text editing.

Specialized Capability Optimization & Multi-Teacher Distillation

Real e-commerce tasks demand simultaneous instruction following, product detail preservation, accurate text rendering, and user-attractive output — objectives with different optimization difficulty and convergence speed. KwaiMind adopts a training scheme combining specialized capability optimization with multi-teacher distillation (DiffusionOPD). On top of general editing, the model reinforces visual quality, product consistency, text accuracy, and commercial click appeal. DiffusionOPD, an online distillation strategy, guides the student along its own denoising trajectory, enabling capability transfer closer to the student's actual generation process. This integrates general editing, text handling, product detail preservation, and user preference into one model, so a single edit can replace background while keeping packaging details and text, and produce a more preference-aligned display.

Ecom-Bench: Systematic E-Commerce Task Definition & Evaluation

Ecom-Bench comprises 11 tasks (100 cases each, 1,100 total) covering model wearing (virtual try-on, garment detail, pose adjustment), poster tasks (background replacement, product extraction, selling-point display), and text tasks (text editing, slogan removal). Evaluation uses 6 general dimensions (instruction following, fusion naturalness, physics/lighting plausibility, non-edited region preservation, image quality, overall aesthetics) and 8 e-commerce dimensions (product identity, text accuracy, fabric texture, model/pose, garment fit, cutout edge, selling-point conveyance, product display completeness). Each task selects 2 general and 2 e-commerce dimensions per its needs. Scores are 1–5 integers per dimension; case composite score uses geometric mean to penalize critical defects (e.g., text errors, product identity shift). Task score averages 100 cases; overall score equally weights 11 tasks.

In the reported evaluation scope, KwaiMind ranks first among open-source models on Ecom-Bench visual total score. An Ecom-CTR metric ranks model outputs by predicted click preference; KwaiMind leads in cumulative top-3 counts, reflecting its balance of material usability and user preference.

Business Validation: Online A/B Test

An online A/B test on product main-image material selection showed KwaiMind delivers a ~2.44% relative click-through rate improvement. Offline, on randomly sampled product images, the proportion of KwaiMind-generated images with higher predicted CTR than the original rose from 12.16% (pre-training) to 37.41%; against unoptimized baseline generations, KwaiMind achieves a 91.89% predicted CTR win rate.

Conclusion & Future Work

KwaiMind demonstrates a path for image editing models into real e-commerce workflows. The unified editing foundation and accompanying evaluation system provide a solid base for e-commerce applications and custom pipelines. For merchants and creators, consolidating these capabilities reduces iterative production costs and bridges creative ideas with usable assets. Future work will expand task coverage, improve specialized reward model robustness, and enhance compositional and multi-reference editing. The project and technical report are publicly available: https://github.com/KwaiMmu/KwaiMind and https://arxiv.org/abs/2609.26375.