Kuaishou open-sources KwaiMind e-commerce image editing model, lifting real CTR by 2.44% in A/B test
机器之心 · wechat · 2026-09-25
Kuaishou released KwaiMind, an image editing foundation model built for real e-commerce content production, combining general editing with domain-specific capabilities.
Results:
- Top overall score among open-source models on multiple general editing benchmarks; #1 visual score on its own Ecom-Bench (11 tasks, 1,100 cases, 14 evaluation dimensions)
- Live A/B test on product hero images: 2.44% relative CTR lift
- Offline: share of generated images beating originals on predicted CTR rose from 12.16% to 37.41%
Tech: A four-agent data engine (filter/generation/caption/coordinator) curates 1.8M training pairs from 26.2M candidates; DiffusionOPD online multi-teacher distillation merges general editing, text accuracy, product consistency and click appeal into one model. Code and technical report are open-sourced.
More from Multimodal
- FineVision, the 17M-image open VLM dataset from 200+ sources, accepted to NeurIPS — andimarafioti · 2026-09-25
- MiniMax H3 seems overtrained on smiles: 'bored caterpillar' video prompt keeps breaking immersion — episodex86 · 2026-09-25
- Pose Blueprint: A Browser-Based 3D Pose Editor for ComfyUI and ControlNet — OkConfusion6667 · 2026-09-25
- Reddit User Explores AI Art With Only Steps, CFG and Denoise Tweaks, No LoRAs — Extreme_Nice · 2026-09-25
- A sub-$20 LoRA makes Qwen-Image 2.1 rotate transparent objects with a prompt — ben_burtenshaw · 2026-09-25
- Lingbot World v2 runs at 60 FPS, hinting world models could reshape game dev — bingxu_ · 2026-09-25