Baidu releases ERNIE-Image-Turbo, a distilled text-to-image model generating in 8 inference steps
Baidu has released ERNIE-Image-Turbo, a distilled text-to-image diffusion transformer that generates images in 8 inference steps. The model runs on consumer GPUs with 24GB VRAM and supports resolutions up to 1376×768, with claimed strengths in text rendering and structured generation tasks.
Baidu releases ERNIE-Image-Turbo, a distilled text-to-image model generating in 8 inference steps
Baidu has released ERNIE-Image-Turbo, a distilled version of its ERNIE-Image text-to-image model that generates images in 8 inference steps. The model is built on a single-stream Diffusion Transformer (DiT) architecture and runs on consumer GPUs with 24GB VRAM.
Technical specifications
ERNIE-Image-Turbo supports multiple resolutions: 1024×1024, 848×1264, 1264×848, 768×1376, 896×1200, 1376×768, and 1200×896. The model uses a guidance scale of 1.0 and operates with bfloat16 precision. Pricing has not been disclosed.
The distillation process used Distribution Matching Distillation (DMD) and reinforcement learning to reduce the 50-step inference requirement of the base ERNIE-Image model to 8 steps while maintaining generation quality, according to Baidu.
Benchmark performance
On GENEval, ERNIE-Image-Turbo with prompt enhancement scored 0.8510 overall, compared to 0.8728 for the base ERNIE-Image model and 0.8481 for FLUX.2-klein-9B. The model achieved 0.9938 on single object detection and 0.8375 on counting tasks.
For text rendering measured on LongTextBench, ERNIE-Image-Turbo scored 0.9655 average across English and Chinese benchmarks, trailing Seedream 4.5 (0.9882) and the base ERNIE-Image model (0.9733) but outperforming FLUX.2-klein-9B (0.5413).
On the OneIG-EN benchmark measuring alignment, text, reasoning, style, and diversity, ERNIE-Image-Turbo scored 0.5656 overall. Nano Banana 2.0 led with 0.5780, while the base ERNIE-Image achieved 0.5750.
Implementation details
The model is available through Hugging Face's diffusers library and SGLang for deployment. Baidu states the model is designed for "posters, comics, multi-panel layouts, and other content creation tasks" requiring text rendering and structured generation.
Two versions are available: ERNIE-Image-Turbo with and without prompt enhancement (PE). The PE version generally shows higher benchmark scores across most metrics.
What this means
ERNIE-Image-Turbo represents Baidu's entry into fast text-to-image generation, prioritizing deployment efficiency over maximum quality. The 8-step generation and 24GB VRAM requirement make it accessible for consumer hardware, though benchmark scores indicate trade-offs compared to the base model. The focus on text rendering and structured layouts positions it for specific use cases like poster and comic generation rather than general-purpose image synthesis. Whether the speed gains justify the quality reduction will depend on application requirements.
Related Articles
Black Forest Labs releases FLUX 3 Image Klein, a 5.6B-parameter open-weights image model
FLUX 3 Image Klein is a 5.6B-parameter open-weights image model covering text-to-image, multi-reference generation and image editing. Its developer claims it leads its own blind evaluations in all three tasks. Pricing, license terms and benchmark data were not disclosed in the announcement.
Moondream 3 Preview lists on OpenRouter: 9B MoE vision model, 2B active, 32K context, $0.30/$2.50
Moondream 3 Preview is a vision-language model with 9B total and 2B active parameters (mixture-of-experts), a 32K-token context window and an optional reasoning mode. OpenRouter lists it at $0.30 per 1M input tokens and $2.50 per 1M output tokens, served by Prime Intellect.
Moondream 3.1: 9B MoE vision model with 2B active parameters, 32K context, $0.30/$2.50 per 1M tokens
Moondream has released Moondream 3.1, a mixture-of-experts vision-language model with 9B total and 2B active parameters and a 32K-token context window. It is listed on OpenRouter at $0.30 per 1M input tokens and $2.50 per 1M output tokens, with an optional reasoning mode.
IBM Granite Switch 4.1 8B and 30B fold chat, decisions and speech-to-text into one endpoint
IBM has released Granite-Switch-4.1-8B and Granite-Switch-4.1-30B, models that bundle a Decision adapter, a Guardian safety adapter and a speech-to-text model behind a single endpoint. IBM says the combined model uses roughly the same GPU VRAM as the base model alone. Benchmark claims are IBM's own, and context window and pricing were not disclosed.
Comments
Loading...