

ERNIE Image API 将百度开源权重的 8B Diffusion Transformer 引入你的技术栈,由 ERNIE-Image Team 基于 Apache 2.0 发布。它在 LongTextBench 上以 0.9733 的成绩位居榜首,能让海报标题和漫画对白气泡保持清晰可读;同时,蒸馏版 Turbo 变体将推理从 50 步缩短到 8 步。Atlas Cloud 通过一个兼容 OpenAI 的端点提供服务,并采用透明的按量付费定价。立即开始构建。
Atlas Cloud 为您提供最新的行业领先创意模型。
最低成本
| Modality | Description |
|---|---|
| ERNIE Image API (Text To Image) | Where the Turbo endpoint prioritizes throughput, the standard ERNIE Image API leans toward maximum output fidelity on the same text-to-image task. It fits final production work such as posters, editorial graphics, and commercial layouts, where getting every detail right outweighs turnaround time. |
| ERNIE Image Turbo API (Text To Image) | Turn a single text prompt into as many as ten images per request across seven aspect ratios, from square 1024 pixels up to 1376 pixels on the long edge. Tuned for low latency, it defaults to eight inference steps and ships a built-in Prompt Enhancer that expands terse prompts before generation. Reach for it when rapid iteration, real-time previews, and high-volume batch runs matter more than squeezing out the last increment of quality. |
From industry-leading text rendering and structured multi-panel layouts to native bilingual prompting, a default prompt enhancer, seven output dimensions, and reproducible Turbo batches, the ERNIE Image API turns precise instructions into production-ready imagery.

A leading LongTextBench score of 0.9733 lets the model render legible, correctly spelled text straight into generated images. Comic speech bubbles, poster headlines, infographic labels, and UI mockup copy all stay sharp and readable.

Generation, edit, composite, and upscale primitives work alongside a grasp of grid-based spatial relationships. Together they yield coherent multi-panel sequences and formatted designs that designers can drive through one centralized pipeline.

Both English and Chinese prompts run natively through the same encoder pipeline, capturing idiomatic phrasing in either language. This dual fluency supports authentic visual storytelling for global campaigns and localized content alike.

Enabled by default, a lightweight Prompt Enhancer rewrites short inputs into richer, structured descriptions before they reach the diffusion backbone. Toggle it off per request whenever literal control over exact wording matters more.

Seven native output sizes span a square 1024x1024, landscape framings up to 1376x768, and portrait shapes down to 768x1376. Each ratio is generated directly, so framing stays intact across every format.

Need volume without the wait? Turbo mode runs as few as 8 inference steps and returns up to 10 images per request, while an explicit seed keeps every result reproducible.
Feed the exact same brief to the flagship ERNIE Image model, a popular rival, and its faster sibling, then judge how each one renders typography, layout, and light side by side.
Top-down flatlay still life, camera locked perfectly overhead looking straight down onto a weathered pale-elm apothecary counter of a traditional Chinese herbal-tea dispensary. Hard directional late-morning window light rakes in low from the right, the true protagonist of the frame — casting long, crisp, elongated shadows that stretch leftward across the raw wood grain and act as leading lines. On the dense right side, tightly clustered clear glass jars glow as the sun passes through them: translucent dried chrysanthemum buds, red goji berries, curled amber tangerine peel (chenpi), and deep crimson dried roselle petals catching the light. A small oxidized brass hand-balance scale with matte patina, a worn stone mortar and pestle dusted with fine powder, and coarse-fibered handwritten paper prescription slips inscribed with neat brush-calligraphy Chinese characters in traditional kaishu ("甘草三钱", "桂花蜜"), edges frayed and fibrous. Caught mid-moment: a toppled pewter canister on its side, its mouth open, several goji berries still rolling and scattering outward, each casting its own thin needle-long shadow. Composition breathes through density-and-void — the packed cluster on the right balanced against a broad expanse of empty bare-wood negative space on the left. Monochromatic warm palette throughout — amber, tangerine-orange, aged brass gold — broken by a single note of dark roselle red. Textures must hold up to magnification: the brittle thinness of dried petals, the dull oxidized brass, the ragged paper fiber edges, the grain of loose powder. Natural directional light, no artificial glow, clean crisp shadows, realistic material rendering, restrained and elegant, macro-detailed food-and-herb still-life photography, shot with an 85mm lens, wide horizontal landscape framing, wide 16:9 aspect ratio, full-bleed.

Generated with Baidu ERNIE Image Turbo on Atlas Cloud

Generated with Qwen Image 2.0 on Atlas Cloud

Generated with Baidu ERNIE Image Turbo on Atlas Cloud
A three-panel horizontal manga strip following a teenage inventor girl in her cluttered attic workshop. In the first panel she sketches a small flying machine by warm lamplight, in the second the contraption sputters and lifts off mid-air scattering bolts, in the third she throws both fists up grinning in triumph. Clean bilingual speech bubbles carry crisp English and Japanese lettering, drawn with confident ink linework and screentone shading, warm amber lamp glow balanced against cool workshop shadows. Character design stays consistent across all three panels, gestures stay expressive, and the story reads left to right with clear sequential flow. Vibrant cel-shaded anime illustration style with bold clean outlines. Wide 16:9 aspect ratio, full-bleed.

Generated with Baidu ERNIE Image Turbo on Atlas Cloud

Generated with Qwen Image 2.0 on Atlas Cloud

Generated with Baidu ERNIE Image Turbo on Atlas Cloud
从文字精准的海报和多格漫画,到双语营销活动、商品目录、界面原型和带标注的信息图,ERNIE Image API 能将精确提示词转化为版式准确的视觉内容,覆盖各类内容生产流程。
凭借模型领先的文字准确性,清晰可读的标题、价格和产品文案可直接渲染到营销海报和横幅中。营销团队无需单独排版步骤,即可直接交付可印刷成品素材。
由于模型理解基于网格的版式和多格结构,它可以渲染连贯的漫画页面,并将对白放入对话气泡中。独立创作者和工作室无需手工重绘每一帧,就能快速起草完整分镜。
原生支持英文和中文提示词,意味着同一套流程即可为两个市场生成符合品牌调性的视觉内容,并在不同文字体系中正确渲染文本。全球团队无需为每种语言单独搭建设计流程,也能完成创意本地化。
通过一次 API 调用,即可为完整商品目录生成生活方式场景、产品模型图和促销图片。Turbo 版本将推理压缩到 eight steps,因此高流量店铺可在数分钟内刷新整套商品目录。
需要用于提案的真实感界面吗?模型可以渲染应用界面和网站原型图,并生成可读的标签、按钮和正文文案,让产品团队在构建任何组件之前,就能获得可用于演示的原型。
强大的指令遵循能力,可在一次生成中将图像与标注清晰的示意图、图表和注释结合起来。教育工作者和分析师可以把密集的源材料转化为说明型图形,并在任何显示尺寸下保持清晰可读。
See where ERNIE Image lands next to other open and proprietary generators across developer origin, access model, bilingual text rendering, and per-image cost.
| Model | Developer | Access Model | Bilingual Text Rendering (EN + ZH) | Price (per image) |
|---|---|---|---|---|
| ERNIE-Image | Baidu (ERNIE-Image Team) | Open weights, Apache 2.0 | Industry-leading, LongTextBench 0.9733 | Pay-as-you-go |
| ERNIE-Image Turbo | Baidu (ERNIE-Image Team) | Open weights, Apache 2.0 | Retained through DMD-distilled 8-step inference | Pay-as-you-go |
| Qwen Image 2.0 | Alibaba (Tongyi) | Open weights, Apache 2.0 | Strong across 1K-token typography layouts | $0.035 |
| Z-Image Turbo | Alibaba (Tongyi Lab) | Open weights, Apache 2.0 | Handles complex Chinese signage alongside English | $0.005 |
| Seedream v4.5 | ByteDance | Proprietary | Designer-level rendering at native 4K | $0.04 |
几分钟即可上手 — 按照以下简单步骤,通过 Atlas Cloud 平台集成和部署模型。
在 atlascloud.ai 注册并完成验证。新用户可获得免费额度,用于探索平台和测试模型。
将先进的 ERNIE Image API for Readable Text in Images 模型与 Atlas Cloud 的 GPU 加速平台相结合,提供无与伦比的性能、可扩展性和开发体验。
低延迟:
GPU 优化推理,实现实时响应。
统一 API:
一次集成,畅用 ERNIE Image API for Readable Text in Images、GPT、Gemini 和 DeepSeek。
透明定价:
按 Token 计费,支持 Serverless 模式。
开发者体验:
SDK、数据分析、微调工具和模板一应俱全。
可靠性:
99.99% 可用性、RBAC 权限控制、合规日志。
安全与合规:
SOC 2 Type II 认证、HIPAA 合规、美国数据主权。
The ERNIE Image API gives developers programmatic access to Baidu's open-weight text-to-image model, an 8B single-stream Diffusion Transformer paired with a Prompt Enhancer that expands short prompts into richer, more structured descriptions. On Atlas Cloud you reach it through one OpenAI-compatible endpoint with pay-as-you-go pricing and Day-0 access.
Its standout strength is legible in-image text. The model scores 0.9733 on LongTextBench in English, the top result among open-weight models, which makes it dependable for posters, comic speech bubbles, infographics, and UI mockups where every character has to be spelled correctly.
Both variants share the same 8B architecture but trade quality against speed. The Standard model runs 50 inference steps at guidance scale 4.0 for maximum fidelity on final assets, while the Turbo variant is distilled with DMD and reinforcement learning down to roughly 8 steps for rapid, high-volume generation.
Yes. Prompts are supported in English, Chinese, and Japanese through the same encoder, and text stays reliable across scripts, scoring 0.9661 on the Chinese LongTextBench. Where several competing models degrade sharply on Chinese characters, this one keeps Simplified, Traditional, and mixed bilingual copy clean.
The Turbo endpoint accepts seven preset sizes through a single size parameter, ranging from a 1024x1024 square to 1376x768 landscape and 768x1376 portrait formats. You can also request up to ten images per call, fix a seed for reproducible results, and toggle the built-in Prompt Enhancer with the use_pe flag.
Getting started takes a single API key. Sign up on Atlas Cloud, point your existing OpenAI-compatible client at the endpoint, and send a prompt with an optional size and seed to receive image URLs in the response. Billing is pay-as-you-go per call with Day-0 access to the model.
In published benchmarks the model outperforms comparable open releases such as FLUX.2-klein-9B, scoring 0.8856 against 0.8481 on GenEval overall. Its widest lead is in text rendering, where FLUX.2 collapses to 0.2183 on Chinese while ERNIE Image holds above 0.96. For workloads built around readable in-image text and structured layouts, it is currently the strongest open-weight choice.
Yes. ERNIE Image is released under the Apache 2.0 license, which permits commercial use, modification, and redistribution. Generated images can go into advertising, merchandise, publications, and other commercial products without license friction.
Join the Discord community for the latest model updates, prompts, and support.