WeLM-Image unified image generation model integrates multimodal understanding
Original title: WeLM-Image: A Unified Image Generation Model towards Visual Intelligence
WeLM-Image combines text-to-image generation, instruction-guided editing and multi-reference synthesis on a single GPU with MoT backbone and MoE modules, ranking first on PRISM-Bench and Qwen-Image-Bench.
WeLM-Image is a unified foundation model processing text, reference images and target latents as one interleaved sequence. A Mixture-of-Transformers backbone with shared attention and independent parameters limits optimization interference. Each path includes Mixture-of-Experts modules providing sparse capacity. A knowledge-grounded visual vocabulary and large-scale clustering balance semantic coverage. Bilingual annotations and editing supervision support training up to 1.5K resolution. Separate optimization with rubric and checklist rewards followed by distillation yields WeLM-Image-Turbo, generating 1024x1024 images in under 5 seconds on a single H20 GPU.
- 31.7B stored parameters, approximately 5.4B activated per target latent token.
- 1536-resolution model ranks first on PRISM-Bench, Qwen-Image-Bench and UnicBench.
- MeanFlow-based few-step distillation enables 1024x1024 generation in 4.98 s on single NVIDIA H20 GPU.
From the original article
1 min read at huggingface.co
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.
By Canmiao Fu, Chen Li, Guohui Zhang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.