VisCPM

Why we included this project

VisCPM gives you one model family that does both halves of vision-language work: chatting about images and generating images from text, in Chinese and English. It builds on the 10B-parameter CPM-Bee language model, with a visual encoder for multimodal conversation and a diffusion UNet for text-to-image, all sharing a single backbone. The clever bit is that training only on English multimodal data transfers well to Chinese, so Chinese-speaking teams don't need to assemble large native paired datasets. The repo includes reproducible training recipes, low-resource inference (around 5GB VRAM for the chat model), fine-tuning support, and a local web demo, which makes it practical for research replication or product integration. If you want a more current checkpoint, the project points to successors like OmniLMM and MiniCPM-V.