A new image model (ERNIE-Image-8b) from Baidu will be released soon.

Baidu's upcoming ERNIE-Image-8B image generation model has been revealed through code leaks ahead of its official announcement. Technical documentation and integration code have surfaced on multiple GitHub repositories, including a pull request for ComfyUI integration (#13369) and another for the Hugging Face Diffusers library (#13432). These leaks confirm the model's 8-billion parameter scale and reveal implementation details through the official Diffusers pipeline documentation.

The model appears to come in two variants: a standard ERNIE-Image model and a faster ERNIE-Image-Turbo version, though the Turbo model's Hugging Face page currently returns a 404 error. The leaks show that developers are already working to integrate ERNIE-Image-8B into popular AI image generation workflows, suggesting it will be compatible with existing tools like ComfyUI's node-based interface. This positions Baidu's model as a direct competitor in the rapidly evolving text-to-image space, potentially challenging established players like Stability AI's SD3 and Midjourney.

The timing of these leaks suggests an imminent official release from Baidu, which has been expanding its ERNIE (Enhanced Representation through kNowledge IntEgration) AI family beyond language models. The 8-billion parameter count places it in the mid-range of current image models, potentially offering a balance between quality and computational efficiency. The Turbo variant's existence indicates Baidu is addressing the speed concerns that often plague large diffusion models, which could make it more practical for real-time applications.