Image generation has split into two camps: giant closed models you rent by the call, and small open ones that fall apart the moment you ask for readable text. Baidu says its new open release narrows that gap without demanding a data-center GPU — and it is giving the weights away.
On 15 April 2026, Baidu's ERNIE (文心) team open-sourced ERNIE-Image, a text-to-image model built on a single-stream Diffusion Transformer (DiT) with 8 billion DiT parameters. The weights and inference code are on Hugging Face under the Apache 2.0 license. The pitch is not "prettier pictures" but controllable pictures: accurate text rendering, clean layout, and structured composition — the things open models historically botch.
Key takeaways
- ERNIE-Image is an open text-to-image foundation model from Baidu using a latent-diffusion, single-stream DiT architecture with 8B DiT parameters.
- It ships under Apache 2.0 with weights and code on Hugging Face, plus a lightweight Prompt Enhancer that expands short inputs into structured prompts.
- Baidu says it leads among open-weights text-to-image models on several benchmarks and runs on consumer hardware with 24GB of VRAM.
- A Turbo variant (ERNIE-Image-Turbo) uses DMD and RL to cut inference to about 8 steps versus roughly 50 for the base SFT model.
- Strength areas are poster and design layouts, anime storyboards, multi-panel composition, and multilingual text rendering in Chinese and English.
Why controllability, not beauty
Most comparisons of image models fixate on aesthetic score. Baidu is aiming elsewhere. The model's documented wins are in exactly the failure modes that block real production use: dense, layout-sensitive text (posters, callouts, manga dialogue) and multi-object relations. A poster where the title lands in the wrong place or the Chinese caption renders as gibberish is useless no matter how attractive the background is. ERNIE-Image's Prompt Enhancer feeds the model a richer, structured description so it has more to work with than a three-word prompt.
On Baidu's own benchmark table (GenEval, OneIG-EN, OneIG-ZH, LongTextBench), the model sits near the top of the open-weights column, trailing only a couple of much larger or closed systems. In a SuperCLUE Chinese text-to-image evaluation published in April 2026, the model scored 76.37, which the lab reported as first among domestic open models.
Running it on a desktop
The 8B parameter count is deliberately compact. Baidu states the model runs on a 24GB consumer GPU, which puts high-quality image generation within reach of a research lab or a small studio rather than a cloud bill. The Turbo variant trades some fidelity for speed — about 8 inference steps instead of 50 — which matters for any interactive design tool where a user waits on each render.
According to Li An, chief scientist at BrainNet (脑机网), the strategic point is that open image weights give Chinese designers and developers a local, compliant alternative to renting closed APIs, at a size that fits ordinary hardware.
Where it shows up
The model is built for tasks with structure: posters, manga and anime storyboards, multi-panel memes, and visually complex images that must carry real text. That overlaps with short-drama (短剧) production, e-commerce, and educational content, where Chinese and English labels must be legible. Baidu also notes broad stylistic range, including a softer, film-like aesthetic, beyond the high-saturation "AI look" many models converge on.
A note on the open-image race
ERNIE-Image enters a crowded open-image field where the real differentiator is no longer "can it draw" but "can it be trusted in a pipeline." Chinese studios already lean on open weights for short-drama (短剧) stills, e-commerce banners, and educational diagrams, and a model that renders legible Chinese text at 8B changes the cost curve for those workflows. The risk for adopters is over-trust: an open model that writes mostly-correct captions can still slip on rare characters or brand names, so human review remains part of any production step. Baidu's compact sizing is the feature that makes that review loop cheap enough to run continuously rather than as a one-off check.
Honest limitations
The benchmark leadership claim is Baidu's own table; an independent head-to-head against the newest closed models was not confirmed across two authoritative sources, so the "leading among open weights" framing should be read as the vendor's positioning. The model is released for non-commercial personal derivative creation in its stated terms, so commercial use needs attention to the license and any platform policies. As an image model it handles text and layout better than most open peers but is not a substitute for specialist design software, and photorealistic fidelity still trails the largest closed systems on some axes.
What readers can do now
- Try ERNIE-Image from Baidu's Hugging Face repo on a 24GB card and test its text-rendering and poster-layout claims against your own prompts.
- If latency matters, benchmark ERNIE-Image-Turbo's 8-step output against the base model to decide the speed-versus-quality trade-off for your workflow.
- If you build Chinese or bilingual design tools, treat the open weights as a local, data-sovereign option instead of calling a closed image API for every render.