WeLM-Image: A Unified Image Generation Model towards Visual Intelligence
Canmiao Fu, Chen Li, Guohui Zhang, Hao Lu, Hongyang Wei, Huadong Wu, Huaisong Zhang, Hubery Yin and 8 others
WeChat Vision, Tencent Inc.
Published November 5, 2026 · on the feed November 6, 2026 · ▲ 48 · arXiv 2611.07392
The abstract
We present WeLM-Image, a unified image foundation model that integrates multimodal understanding into text-to-image generation, instruction-guided editing, and multi-reference synthesis. Designed for single-GPU deployment in WeChat's large-scale consumer applications, the model processes text, reference images, and target latents as one interleaved sequence. A Mixture-of-Transformers (MoT) backbone combines shared attention with independent understanding and generation parameters to limit optimization interference. Equipping each path with Mixture-of-Experts (MoE) modules provides sparse capacity, with 31.7B stored parameters and approximately 5.4B generation parameters activated per target latent token. A generation-oriented data engine combines a knowledge-grounded visual vocabulary with large-scale visual clustering to balance semantic coverage and editing tasks, while training-time evaluation guides iterative data refinement. Bilingual annotations and verified editing supervision support progressive training up to native 1.5K resolution. We optimize generation and editing separately with rubric- and checklist-based rewards, then consolidate the specialized policies through noise-aware on-policy distillation. In the reported overall comparisons, the 1536-resolution model ranks first among the evaluated open-source models on PRISM-Bench, Qwen-Image-Bench, and UnicBench. MeanFlow-based few-step distillation yields WeLM-Image-Turbo, which generates a 1024times1024 image in 4.98\,s on a single NVIDIA H20 GPU using eight guidance-free steps.
The Brief
2:18A short podcast summary of the paper. Written and voiced by AI.
Transcript
WeLM-Image. A Unified Image Generation Model towards Visual Intelligence. By Canmiao Fu and fifteen others, from WeChat Vision at Tencent. Published on the fifth of November.
WeLM-Image is meant to do several jobs in one model: text-to-image generation, editing guided by instructions, and synthesis from multiple reference images, all with multimodal understanding built in. It also has a hard deployment constraint. It is designed to run on a single GPU inside WeChat's large-scale consumer applications. Capability and footprint pull in opposite directions, and the design has to answer both.
The model treats text, reference images and target latents as one interleaved sequence. Its backbone is a Mixture-of-Transformers, which shares attention across the whole sequence while keeping separate parameters for understanding and for generation, so the two objectives interfere less during optimization. Each of those paths is fitted with Mixture-of-Experts modules for sparse capacity. The result stores thirty-one point seven billion parameters, but activates only about five point four billion generation parameters for each target latent token.
Data comes from a generation-oriented engine. It pairs a knowledge-grounded visual vocabulary with large-scale visual clustering, to balance semantic coverage against editing tasks, and evaluation during training guides repeated rounds of data refinement. Bilingual annotations and verified editing supervision support progressive training, all the way up to native one point five K resolution.
Generation and editing are then optimized separately, with rewards built from rubrics and checklists. The two specialized policies are merged back into one through noise-aware on-policy distillation.
In the reported overall comparisons, the model at one thousand five hundred and thirty-six resolution ranks first among the open-source models evaluated on three benchmarks: PRISM-Bench, Qwen-Image-Bench and UnicBench.
A faster variant, WeLM-Image-Turbo, comes from few-step distillation based on MeanFlow. Using eight guidance-free steps on a single NVIDIA H20 GPU, it produces a one thousand and twenty-four by one thousand and twenty-four image in four point nine eight seconds.
Read as a whole, WeLM-Image is an exercise in fitting breadth into a budget. Sparse experts keep the active parameter count far below the stored one, separate paths keep understanding and generation from interfering, and a distilled Turbo variant brings generation down to a handful of steps.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is three image jobs, one model, and a single GPU to run it on.
You are building one image model for a messaging app used at huge scale. It must generate images from text, edit images on instruction, and combine several reference images.
The road ahead
- 1 · The problem ~6 min One image model for a huge messaging app must generate, edit and combine references, yet run cheaply on a single GPU.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Put all three tasks in one interleaved sequence, give understanding and generation separate sparse experts joined by shared attention, and train generation and editing apart before merging.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Curate balanced data, pretrain up to 1.5K resolution, reinforce generation and editing separately, merge them by distillation, then distill a fast Turbo model.
- 6 · The experiment ~6 min You’ll guess: Among the evaluated open-source models, where did the 1536-resolution WeLM-Image rank overall on PRISM-Bench, Qwen-Image-Bench and UnicBench?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.