KwaiMind Technical Report
Junlong Wu, Zijun Li, Yuting Hu, Jia Sun, Pengcheng Wei, Yimin Zhou, Honglie Wang, Huaiqing Wang, Dewen Fan, Fei Zuo, Haixuan Gao, Lihui Peng, Tingxuan She, Yuqing Li, Boheng Zhang, Fan Yang, Wenwu Ou
cs.CV
2026-09-22
Kuaishou's KwaiMind leads open-source ImgEdit (4.15) and Ecom-Bench (3.67); CTR RL lifts beat-original share from 12.16% to 37.41% and live selected CTR about 2.44%.
Instruction-based editors already add, remove, restyle, and recompose. Storefront images add three constraints those benchmarks barely score: the SKU cannot drift in color, texture, or branding; promo text has to be spelled and placed correctly; and the frame has to attract clicks. Imitation learning on public edit pairs does not optimize any of those.
KwaiMind is Kuaishou's attempt to move data, rewards, and evaluation onto that commercial spec. The backbone follows Qwen-Image-Edit-2511: a dual-stream MMDiT. The data engine keeps about 1.8 million pairs. Ecom-Bench covers 11 merchant tasks.
A coordinator agent runs Filter, Generation, and Caption as a state machine, with humans only on uncertain and out-of-distribution cases. 26.2 million raw pairs (ScaleEdit-12M, X2Edit, AnyEdit, plus 8.0 million internal KwaiData) are filtered, repaired, and gap-filled down to about 1.8 million. Captions come in three styles (detailed, short, colloquial) and are reverse-audited. SFT keeps 234.9K pairs: 115.0K general and 119.9K e-commerce, spanning try-on, garment detail, display, wearing, pose change, background swap, outpaint, cutout, text edit, tagline removal, and selling-point posters. Multi-reference tasks feed both a model photo and a product photo.
Training has four stages: continued pre-training above 720p with bilingual instruction sampling; SFT; mixed offline and online Diffusion-DPO; then DiffusionNFT online RL on the forward process. Gemini 3.1 Pro Preview grades general quality on image, instruction, and aesthetics (1-5). Three specialist rewards cover what the judge misses.
The CTR model trains on same-product pairs with more than 2,000 impressions and a stable 7/14/30-day order, first as 10-bucket classification then gap-weighted ranking. In-house pairwise accuracy is about 70%. On public CreativeRanking, a vision-only variant reaches 57.1 PairAcc versus CAIG 56.2 and GPT-4o 50.2. A zero-shot VLM CTR judge sits near chance. The OCR reward mixes span placement with glyph-level recognizer posteriors so missing strokes still hurt. The consistency reward scores product identity and target-part fidelity on a conservative 1-4 rubric.
Each objective gets its own NFT teacher. DiffusionOPD then matches student transition means to those teachers on the student's own trajectories, so heterogeneous rewards are not trained as one soup and sequential RL does not wipe earlier skills.
KwaiMind leads the open-source group on four general suites (Gemini 3.1 Pro Preview on ImgEdit/GEdit; native protocol on REDEdit):
| Bench | KwaiMind | Open-source #2 | Closed best |
| ImgEdit | 4.15 | FireRed-Edit 4.11 | Seedream 4.27 |
| GEdit | 6.29 | Qwen-Edit-2511 5.97 | GPT-Image-2 7.10 |
| REDEdit EN | 3.75 | FireRed-Edit 3.67 | GPT-Image-2 4.17 |
| REDEdit CN | 3.73 | FireRed-Edit 3.67 | GPT-Image-2 4.22 |
Against its Qwen-Edit-2511 init, ImgEdit moves from 3.83 to 4.15 and GEdit from 5.97 to 6.29. The closed-source gap remains, especially GEdit at 7.10.
Ecom-Bench uses 100 images per task and a geometric mean of four dimensions so one critical miss sinks the sample. Visual overall is 3.67, best among open models; GPT-Image-2 is 4.20 and NanoBanana-2 4.01. The open-source lead is concentrated on Clothing Detail 4.25 (FireRed-Edit 1.82), Clothing Display 4.19, and Pose Change 4.44. Virtual Try-On 3.98 trails FireRed-Edit 4.27. Outpaint 2.55 is far behind Flux-2.0 at 4.57; Product Extract 3.69 trails Joy-Edit 3.97. Joy-Edit and LongCat skip the two-image tasks, so their overall scores average nine tasks, not eleven.
On 1,100 CTR rankings across eight models, KwaiMind occupies rank 1/2/3 in 215/164/165 cases (sum 544), ahead of Seedream 525 and GPT-Image-2 431. Offline, the share of generations whose predicted CTR beats the original product image rises from 12.16% to 37.41% after CTR RL. An online A/B that selects main images with this scorer reports about +2.44% live CTR. That experiment is selection, not a generator in the treatment arm.
For merchant creatives, identity, glyphs, and clicks are the production gates. KwaiMind turns those into train-time rewards and distills them into one editor that already wins the open-source tables on general edits and garment tasks. Proprietary visual scores are still clearly higher. The transferable recipe is continued training on a Qwen-Image editor, specialist rewards from logs, then on-policy merge. Generic VLMs are not a CTR judge.
There is no standalone limitations section. The conclusion flags the closed-source visual gap plus weak compositional and multi-reference edits. Train-time and Ecom-Bench judges are both Gemini 3.1 Pro Preview, so family bias is untested. The CTR ranking uses Kuaishou's own reward on everyone's outputs; KwaiMind was optimized for that score, so 544 is not "sells best online." The +2.44% A/B does not report traffic, category mix, or whether generated images were in the treatment set. Outpaint and cutout lag, which matches rewards that never targeted those tasks. A GitHub repo is listed; the report does not say whether full weights are public.