ERNIE Image is Baidu's 8B parameter open-source text-to-image model, released April 15, 2026 under the Apache 2.0 license. Built on a single-stream Diffusion Transformer (DiT) with a lightweight Prompt Enhancer that expands short prompts into detailed visual descriptions before generation. Scores 0.9733 on LongTextBench (text rendering) and 0.8856 on GenEval (instruction following). Handles posters, infographics, UI mockups, comics, multi-panel layouts, and bilingual Chinese/English text with precision. Two variants: SFT (50 steps, max quality) and Turbo (8 steps, 6x faster). Runs on a single 24GB consumer GPU. Available as a ComfyUI node on Floyo.

ERNIE IMAGE WORKFLOWS ON FLOYO

ERNIE Image - Text to Image

ERNIE Image is Baidu's open-source text-to-image model, released April 15, 2026. It is an 8 billion parameter single-stream Diffusion Transformer (DiT) paired with a lightweight Prompt Enhancer. With only 8B DiT parameters, it matches or exceeds larger open-source models across multiple benchmarks. The model is specifically designed for tasks that trip up most image generators: legible in-image text, structured layouts, posters, infographics, comics, and multi-panel compositions.

The Prompt Enhancer is what makes ERNIE Image forgiving for short prompts. Type "an abandoned Victorian mansion overtaken by vines, oil painting style" and the enhancer rewrites it into a detailed visual description with lighting, mood, and composition before the DiT sees it. This compensates for the smaller model size: a smart enhancer plus a focused 8B DiT produces output that competes with 20B+ models.

Text rendering is ERNIE Image's strongest capability. It scores 0.9733 on LongTextBench, the benchmark for dense, long-form, and layout-sensitive text in generated images. Posters with multi-line headlines, infographics with data labels, UI mockups with button text, and menus with item lists all render legibly. Both Chinese and English text render cleanly in the same generation pass.

Two variants ship with the release. The SFT variant runs at guidance scale 4.0 and 50 steps for maximum quality. The Turbo variant uses DMD (Diffusion Model Distillation) and reinforcement learning to compress inference from 50 steps to 8, achieving 6x speed improvement while maintaining high quality output.

On Floyo, ERNIE Image runs through native ComfyUI nodes on H100 NVL GPUs. The workflow includes the Prompt Enhancer toggle, configurable resolution, steps, CFG, and seed. Type a prompt and generate. No model downloads, no local GPU required.

ERNIE Image uses an 8B parameter single-stream Diffusion Transformer with a lightweight Prompt Enhancer (Ministral 3B). Two variants: SFT (50 steps, CFG 4.0, max quality) and Turbo (8 steps, CFG 1.0, 6x faster). Default resolution is 1024x1024. Runs on a single 24GB consumer GPU. Bilingual Chinese/English prompts and in-image text. Apache 2.0 licensed with open weights on HuggingFace.

ERNIE Image's feature set is built around one insight: a smaller model with a smart prompt enhancer can match a larger model that takes raw prompts directly. The 8B DiT focuses on rendering. The Ministral 3B enhancer handles prompt understanding. This separation of concerns is why ERNIE Image punches above its weight class.

Built-in Prompt Enhancer

A lightweight Ministral 3B LLM rewrites your short prompt into a detailed visual description before the DiT sees it. A one-line idea like "cyberpunk street market at night" becomes a full paragraph describing neon colors, rain reflections, vendor stalls, atmospheric haze, and camera angle. This compensates for the 8B model's limitations in complex prompt understanding. Toggle it off when you want your exact wording untouched.

Text Rendering (0.9733 LongTextBench)

The strongest text rendering benchmark score among open-source models at this parameter count. Dense, long-form, and layout-sensitive text renders legibly. Posters with multi-line headlines, infographics with data labels and annotations, UI interfaces with button text, and menus with item lists all come out readable. Both Chinese and English text in the same image.

8B Efficiency

Runs on a single consumer GPU with 24GB VRAM. The model footprint in bfloat16 is about 29.5GB, which fits on RTX 4090 and similar cards with CPU offloading. Despite being significantly smaller than competitors like Qwen-Image (20B) or HunyuanImage 3.0 (80B MoE), it matches or exceeds them on text rendering and instruction following benchmarks.

Turbo Variant (6x Speed)

The Turbo variant uses DMD (Diffusion Model Distillation) and reinforcement learning to compress inference from 50 steps to 8. This produces 6x speed improvement while maintaining high quality. Use SFT for final production assets and Turbo for fast iteration and previews. Both share the same architecture and produce compatible output.

Structured Layouts and Comics

ERNIE Image handles multi-panel layouts, comic pages, and structured compositions that most image models cannot produce coherently. Panel boundaries, text placement, and visual hierarchy are maintained across the full image. This extends the model's use beyond single-scene generation into sequential visual storytelling.

Apache 2.0 License

Fully open source with commercial rights. Weights are on HuggingFace. Day-0 ComfyUI support was added in April 2026. Diffusers and SGLang deployment paths are both documented. Fine-tuning is supported through AI-Toolkit. GGUF weights are available through Unsloth.