AI news ·
WeLM-Image ranks first among evaluated open-source models on three image generation benchmarks
Tencent built a 31.7-billion-parameter image model that runs on a single GPU and tops three major benchmarks. It generates a 1024×1024 image in under 5 seconds on one NVIDIA H20 GPU.
Tencent's WeChat team has built a 31.7-billion-parameter image model that runs on a single GPU and already tops three major benchmarks. The model, called WeLM-Image, is designed for real-world consumer applications inside WeChat, handling text-to-image generation, instruction-guided editing, and multi-reference synthesis within one unified system.
The architecture addresses a practical problem: large image models typically demand multi-GPU setups that are impractical for consumer-scale deployment. WeLM-Image uses a Mixture-of-Transformers backbone that separates understanding and generation parameters, then activates only about 5.4 billion parameters per generated token through Mixture-of-Experts modules. The result is a model that can produce a 1024×1024 image in under five seconds on a single NVIDIA H20 GPU.
How the model handles multiple tasks in one sequence
Unlike systems that bolt separate modules together, WeLM-Image processes text prompts, reference images, and target latents as one interleaved sequence. The Mixture-of-Transformers design uses shared attention layers while keeping independent parameters for understanding and generation, which limits the optimization interference that often degrades performance when combining these capabilities.
A generation-oriented data engine supports this unified approach. It combines a knowledge-grounded visual vocabulary with large-scale visual clustering to balance semantic coverage and editing precision. The team used bilingual annotations and verified editing supervision, with training-time evaluation guiding iterative data refinement. Training progressed up to a native resolution of 1536 pixels.
Benchmark results and few-step distillation
The 1536-resolution version of WeLM-Image ranks first among evaluated open-source models on PRISM-Bench, Qwen-Image-Bench, and UnicBench. These benchmarks test a range of capabilities including prompt following, visual quality, and editing accuracy.
The team also developed WeLM-Image-Turbo using MeanFlow-based few-step distillation. This variant generates a 1024×1024 image in 4.98 seconds on a single H20 GPU using eight guidance-free steps. For professionals working with AI for Creatives Courses, that speed-to-quality ratio changes what's practical in production workflows.
Training with rubric and checklist rewards
The training process optimized generation and editing separately using rubric-based and checklist-based rewards. After specialization, the team consolidated the policies through noise-aware on-policy distillation. This approach preserves editing precision without sacrificing generation quality-a tradeoff that often plagues unified models.
Why this matters for creatives, marketers, and product teams
Single-GPU deployment at this quality level means image generation and editing can run on infrastructure that teams already have, rather than requiring expensive multi-GPU clusters. For marketing teams producing volume visuals, event designers iterating on branded imagery, or product developers prototyping concepts, the 5-second generation time on commodity hardware removes a bottleneck that has limited practical adoption. Writers and PR professionals gain a tool that can produce publication-ready images from text instructions without waiting for design resources. The benchmark results suggest the quality holds up against the best available open models, making the speed and deployment efficiency genuine advantages rather than compromises.