Paper review · November 7, 2026
WeLM-Image: A Unified Image Generation Model Toward Visual Intelligence
WeLM-Image: A Unified Image Generation Model towards Visual Intelligence
WeLM-Image uses a MoT–MoE backbone with separate understanding and generation pathways, together with reinforcement learning and distillation, to handle generation, editing and multi-reference composition in one model. The model can be deployed on a single GPU.
Problem
Image generation and editing models need fine-grained control: understanding long Chinese and English prompts, rendering text accurately, editing while preserving identity, and combining several reference images. Large consumer services such as WeChat also have hard requirements for single-GPU deployment and low cost per request.
Method
Text, reference images and target latents are processed as one interleaved sequence. The model uses a MoT (Mixture-of-Transformers) architecture, in which the understanding (UND) and generation (GEN) pathways have separate parameters and share only attention, and adds a separate MoE to each pathway. It stores 31.7B parameters, and about 5.4B generation parameters are active per target latent token. The VAE was trained separately with reconstruction, semantic alignment and text-focused decoder fine-tuning. The data was built using a knowledge-based visual vocabulary and large-scale clustering, and training was scaled up gradually to 1.5K resolution. Generation and editing were each trained with reinforcement learning using rubric and checklist rewards, then merged into one model through noise-aware on-policy distillation. A Turbo version was also built with MeanFlow-based few-step distillation.
Results
- The authors report that the 1536-resolution model ranked first among the evaluated open-source models in an overall comparison on PRISM-Bench, Qwen-Image-Bench and UnicBench.
- WeLM-Image-Turbo generates a 1024×1024 image in 8 steps without guidance in 4.98 seconds on a single NVIDIA H20 GPU. This is 11.4x faster than the 50-step base model with CFG.
- Turbo completes edits with 1–4 input images in 6.78–13.17 seconds.
- WeLM-Image-VAE reached a PSNR of 34.89 on COCO 512, higher than FLUX.2-VAE (34.61) and Qwen-VAE (33.60). In ImageNet 256 generation experiments, it also had the lowest FID among 16-channel VAEs at 12.21 (Qwen-VAE 15.29) and the highest IS at 97.96.
Limitations
The benchmark rankings compare only the open-source models that were evaluated, and the provided text includes only some of the per-benchmark scores. WeGenEditBench and the poster reconstruction set are internal evaluation sets. Its generation FID is higher than that of the 32-channel FLUX.2-VAE (9.87). Speed was measured on only one GPU type, the H20.
Why it matters
The paper describes in detail how generation, editing and multi-reference composition were combined in one model within the cost limits of a real large-scale service, including the data and training process. It is a useful reference for teams building image generation services.
Key terms
- MoT(Mixture-of-Transformers)
- An architecture that gives each modality or task its own parameter pathway while sharing attention. This reduces interference between them while keeping information exchange.
- MoE
- An approach that activates only some of several expert FFNs for each token. This increases model capacity while keeping the amount of computation low.
- On-policy Distillation
- A knowledge transfer method in which a student model learns to follow the teacher model's predictions on samples the student generated itself.
- MeanFlow
- A few-step generation technique that learns the average velocity of a flow so that images can be generated in only a few sampling steps.
Original
Authors: Canmiao Fu, Chen Li, Guohui Zhang, Hao Lu, Hongyang Wei, Huadong Wu, Huaisong Zhang, Hubery Yin, et al. · Posted 2026-11-05 · License CC BY 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.