HunyuanImage 2.1 image model

HunyuanImage-2.1 is a text-to-image model that makes 2K resolution pictures while keeping costs low.

The core parts are tuned for speed and detail. The high-compression VAE lines up its features with DINOv2 to make training efficient, letting 2K images run at the same cost as 1K images in other models. Dual text encoders capture scene descriptions and text rendering, while the 17B parameter DiT backbone runs both single- and dual-stream flows. RLHF training adds another layer of polish.

Main features. It makes high-quality 2K images with cinematic framing. It supports Chinese and English prompts. It offers flexible aspect ratios. It improves text rendering with ByT5. And it rewrites prompts automatically for richer detail.

How it compares. In tests, HunyuanImage-2.1 scored best among open-source models for prompt following, reaching levels close to closed-source commercial ones like GPT-Image and Seedream-3.0. In head-to-head human evaluations, it outperformed Qwen-Image and came close to Seedream-3.0.

System needs. A CUDA-capable NVIDIA GPU is required, with at least 24 GB GPU memory for 2048x2048 image generation. It runs on Linux.

Usage. The model only makes 2K images. Lower resolutions cause artifacts. Best results come with the full pipeline, including prompt enhancement and the refiner.