What is ERNIE-Image (and Turbo)?
ERNIE-Image is Baidu's first standalone open text-to-image model, released in April 2026 in two checkpoints: ERNIE-Image, a supervised fine-tune that runs at 50 steps, and ERNIE-Image-Turbo, distilled with DMD and reinforcement learning to run in 8. Both use a single-stream diffusion transformer of about 8 billion parameters, a Ministral 3 based text encoder and the FLUX.2 VAE, plus an optional prompt enhancer that rewrites short prompts into detailed descriptions.
The weights are under Apache 2.0, and Baidu says the model runs on consumer GPUs with 24GB of VRAM. It supports seven preset resolutions, from 1024x1024 to 1376x768. Baidu aims it at posters, comics, multi-panel layouts and other images where text and layout must be right.
How much does ERNIE-Image (and Turbo) cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | Apache 2.0 |
| fal (Turbo) | $0.01 per megapixel | |
| fal (ERNIE-Image) | $0.03 per megapixel |
Free weights under Apache 2.0. On fal, Turbo costs $0.01 per megapixel and the 50-step model $0.03 per megapixel.
Is ERNIE-Image (and Turbo) worth it?
ERNIE-Image looks better on benchmarks than it does to human voters. Baidu's own tables put it first among open models on GenEval at 0.8856 and level with Seedream 4.5 on OneIG-EN at 0.5750 with the prompt enhancer on, and its LongText-Bench average of 0.9733 is second only to Seedream 4.5 in the comparison.
Artificial Analysis's preference board disagrees. ERNIE Image Turbo ranks 81st of 108 at 915 Elo and the 50-step ERNIE Image 84th at 909, below Z-Image Turbo at 936 and far below the open leader Qwen-Image-2.1 at 1045. The distilled Turbo scoring higher than the slower model fits Baidu's description of it as tuned for aesthetics.
Its value is the licence and the footprint. Apache 2.0 has no revenue cap, the transformer of about 8 billion parameters runs on a 24GB card, and Turbo costs $0.01 per megapixel on fal, so a 1024x1024 image is about a cent.
The prompt enhancer is not a free win. With it switched on, GenEval falls from 0.8856 to 0.8728 and attribute binding drops, even as OneIG and long-text scores rise, so teams writing many short prompts should test both settings. It is a separate language model, which Baidu suggests serving on its own with vLLM for speed.
Where it earns its place
A four-person webcomic team
Draft multi-panel pages and storyboards with captions
Run Turbo locally in ComfyUI at 8 steps on a 24GB card for drafts and switch to the 50-step model for final panels. Apache 2.0 means no licence fee when the comic sells.
An e-commerce platform's growth team
Produce Chinese and English promotional banners with readable prices at volume
Call fal's Turbo endpoint at $0.01 per megapixel, so 50,000 1024x1024 banners cost about $500 a month. Test with the prompt enhancer on and off, since it helps text and hurts attribute binding.
An ML engineer at a design-software company
Serve an image endpoint inside the product
Deploy with SGLang's OpenAI-style endpoint, or run the prompt enhancer separately on vLLM for speed; NVIDIA also packages the model as a NIM container.
Before you commit
- Setup
- Install Diffusers from source and load ErnieImagePipeline, or run sglang serve with the model path; Baidu says a 24GB GPU is enough.
- Integrations
- Diffusers, SGLang, vLLM for the prompt enhancer, ComfyUI, NVIDIA NIM and hosted endpoints on fal.
- Data and privacy
- Apache 2.0 weights run fully offline; the model card does not describe training data sources or scale.
- Cost at scale
- On fal, 100,000 one-megapixel images cost about $1,000 with Turbo or $3,000 with the 50-step model; self-hosted, one 24GB GPU per worker.
- Limits
- Seven preset resolutions up to 1376x768; text-to-image only; the enhancer can weaken attribute binding; no published speed figures.
What’s good
- Apache 2.0 with no revenue cap
- Runs on a 24GB consumer GPU
- Strong long-text and layout scores in Baidu's tests
- Turbo at $0.01 per megapixel on fal
What’s not
- 81st of 108 on Artificial Analysis's text-to-image board
- Prompt enhancer lowers some benchmark scores
- Seven preset resolutions, none above 1376 pixels
- No editing variant
Best for
Developers who need a commercially unrestricted open model for posters, comics and text-heavy layouts on a single 24GB GPU.
ERNIE-Image (and Turbo) alternatives
ERNIE-Image (and Turbo): common questions
- Is ERNIE-Image (and Turbo) worth it?
- ERNIE-Image sets text and layout well and fits on a 24GB card under Apache 2.0, but arena voters rank it below Z-Image-Turbo. We rate it 3.6 out of 5.
- How much does ERNIE-Image (and Turbo) cost?
- Free weights under Apache 2.0. On fal, Turbo costs $0.01 per megapixel and the 50-step model $0.03 per megapixel.
- What are the downsides of ERNIE-Image (and Turbo)?
- 81st of 108 on Artificial Analysis's text-to-image board; prompt enhancer lowers some benchmark scores; seven preset resolutions, none above 1376 pixels; no editing variant.
- Who is ERNIE-Image (and Turbo) best for?
- Developers who need a commercially unrestricted open model for posters, comics and text-heavy layouts on a single 24GB GPU.
- How much does ERNIE-Image (and Turbo) cost for a team or in production?
- On fal, 100,000 one-megapixel images cost about $1,000 with Turbo or $3,000 with the 50-step model; self-hosted, one 24GB GPU per worker.
- Is ERNIE-Image (and Turbo) safe for company data?
- Apache 2.0 weights run fully offline; the model card does not describe training data sources or scale.
- What are the best ERNIE-Image (and Turbo) alternatives?
- The closest alternatives we have reviewed are Z-Image-Turbo, GLM-Image, Qwen-Image-2.1 and MOSS-Image.
Answers checked Nov 7, 2026.