Baidu releases open 8B ERNIE-Image model
Baidu has released ERNIE-Image, an open-weights 8B text-to-image model that it says beats every other open model at text rendering and structured layouts, and runs on a 24GB consumer GPU.
Baidu's ERNIE-Image team has introduced ERNIE-Image, an open-weights text-to-image model. It is a single-stream Diffusion Transformer (DiT) with 8B parameters, built in a latent diffusion framework. The model comes with a lightweight Prompt Enhancer (PE) that turns short prompts into longer, more structured ones.
Baidu describes the model as built for control as much as for looks. The team says it does well at following complex instructions, rendering dense or layout-sensitive text in Chinese, English and other languages, and producing structured images such as posters, manga storyboards and multi-panel sets. Baidu says it runs on consumer hardware with 24GB of VRAM.
There are two versions. The SFT model typically needs 50 inference steps. The Turbo model, optimized with DMD and reinforcement learning, needs 8 steps.
Baidu reports results on GenEval, OneIG-EN, OneIG-ZH and LongTextBench. On OneIG-EN, ERNIE-Image with the Prompt Enhancer scores 0.5750 overall. That puts it ahead of the open-weights models Z-Image (0.5460) and Qwen-Image (0.5390) and level with the closed models Seedream 4.5 (0.5760) and Nano Banana 2.0 (0.5780). Without the enhancer, the score drops to 0.5537, so much of the gain comes from prompt expansion. Turbo scores 0.5656. All figures are self-reported. The provided material does not include the results for the other three benchmarks.
The post says the model is intended for "non-commercial personal derivative creation." It does not specify the license terms.