WeChat's WeLM-Image paper: MoT-MoE architecture for single-GPU 31.7B image generation
bdsqlsz · x · 2026-11-08
WeLM-Image is WeChat's unified image foundation model integrating multimodal understanding into text-to-image generation, instruction-guided editing, and multi-reference synthesis, designed for single-GPU deployment in WeChat's consumer apps.
Architecture
- Text, reference images, and target latents processed as one interleaved sequence
- Mixture-of-Transformers backbone: shared attention with independent understanding/generation parameters to limit optimization interference
- MoE modules per path: 31.7B stored parameters, 5.4B activated per target latent token
- 27-layer MoT-MoE decoder with MLA projections, unified 3D RoPE, 64 routed experts (Top-6 for understanding, Top-20 for generation)
Data & training
- Generation-oriented data engine: knowledge-grounded visual vocabulary + large-scale visual clustering
- Bilingual annotations, progressive training up to native 1.5K resolution
- Generation and editing optimized separately with rubric/checklist rewards, consolidated via noise-aware on-policy distillation
Weights are not open-sourced. Demo images show style transfer, poster design, pose adjustment, collage, and passerby removal.
Related event: WeChat Team Releases 31.7B Unified Image Model WeLM-Image, Weights Not Open(2 posts)→
More from Multimodal
- AI creator turns Gladiator and The Walking Dead into 1970s-style videos — techhalla · 2026-11-08
- Viggle's Remake agent generates a full 4-minute face-swap MV in one shot — lxfater · 2026-11-08
- AI-Generated Product Concept: Batman Mug and Bat-Wing Coaster Set — Delahuntagram · 2026-11-08
- 'Pixar Is Cooked': Viral AI-Generated Video Sparks Debate Over Entertainment's Future — ZabihullahAtal · 2026-11-08
- thespite teases organic perforated 3D forms after long iteration — thespite · 2026-11-08
- MOSS image model parameter test: Karras & Exponential schedulers break generations, five others nearly identical — Officially_Beck · 2026-11-08