Tencent-Hunyuan/HunyuanImage-2.1

Builds on a dual-encoder architecture combining a multimodal LLM for semantic understanding with a character-aware ByT5 encoder for multilingual text rendering, enabling precise alignment between prompts and visual details. Supports FP8 quantization and CPU offloading to reduce memory footprint to 24GB, and features a two-stage pipeline with a base diffusion transformer and optional refiner model. Integrates with Hugging Face model hub and offers ComfyUI workflow support for deployment flexibility.

671 stars.