Baidu Wenxin Team Open-Sources 8B Parameter Text-to-Image Model ERNIE-Image, Runable on Consumer GPUs

Baidu Wenxin Large Model Team officially open-sourced the text-to-image model ERNIE-Image. With only 8B parameters, this model can generate ultra-realistic and complex images comparable to top commercial models on consumer GPUs with 24GB VRAM.

ERNIE-Image is based on the single-stream Diffusion Transformer architecture and equipped with a lightweight Prompt Enhancer to expand input descriptions. It excels in complex instruction following, text rendering, and structured image generation (e.g., posters, comics), covering multiple styles like realism and anime.

Currently, the model's weights and inference code are open-sourced on Hugging Face (under Apache2.0 license), supporting ComfyUI Workflow, and has launched GGUF quantization scheme in collaboration with Unsloth, lowering deployment barriers.