LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes

Hugging Face Daily Papers Papers
diffusion-transformer image-generation open-source pre-training vision-language muon-optimizer state-of-the-art

Summary

LLaDA-Image presents a unified framework that combines a 6B diffusion transformer with a frozen vision-language module for generating photorealistic images with precise editing, achieving state-of-the-art results among open-source models through efficient training and fast inference.

We introduce LLaDA-Image, a unified framework that pairs a 6B Diffusion Transformer (DiT) trained from scratch with a frozen vision-language understanding module built on the LLaDA2.0-Mini diffusion language model backbone. Instead of relying heavily on paired image-text data from the beginning, we first build a strong visual generative prior through image-only pre-training and mid-training. The generation pipeline comprises 220M samples, 98 of which are real images. For efficient and scalable optimization, we use parameter-free RMSNorm throughout the DiT together with the Muon optimizer. The resulting unified model produces highly photorealistic images while accurately following fine-grained editing instructions. We further distill LLaDA-Image into LLaDA-Image-Turbo, enabling fast inference in 2-4 sampling steps. On Qwen-Image-Bench, LLaDA-Image achieves overall scores of 53.53 and 53.38 on the English and Chinese tracks, respectively, setting a new state-of-the-art among open-source models on both tracks. To support further research on capable and efficient generative models, we release our model weights, training code, and detailed recipes.
Original Article
View Cached Full Text

Cached at: 09/04/26, 03:58 AM

Paper page - LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes

Source: https://huggingface.co/papers/2609.03796 Published on Sep 3

#1 Paper of the day Authors:

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

Abstract

LLaDA-Image unifies a 6B diffusion transformer with a frozen vision-language module, using image-only pre-training and a Muon optimizer to generate photorealistic images with precise editing, and is distilled into a fast 2-4 step variant that achieves state-of-the-art open-source results.

We introduce LLaDA-Image, a unified framework that pairs a 6BDiffusion Transformer(DiT) trained from scratch with a frozenvision-language understandingmodule built on theLLaDA2.0-Minidiffusion language modelbackbone. Instead of relying heavily on paired image-text data from the beginning, we first build a strong visual generative prior throughimage-only pre-trainingand mid-training. The generation pipeline comprises 220M samples, 98 of which are real images. For efficient and scalable optimization, we useparameter-freeRMSNormthroughout theDiTtogether with theMuon optimizer. The resulting unified model produces highly photorealistic images while accurately following fine-grained editing instructions. We further distill LLaDA-Image intoLLaDA-Image-Turbo, enabling fast inference in 2-4 sampling steps. On Qwen-Image-Bench, LLaDA-Image achieves overall scores of 53.53 and 53.38 on the English and Chinese tracks, respectively, setting a new state-of-the-art among open-source models on both tracks. To support further research on capable and efficient generative models, we release our model weights, training code, and detailed recipes.

View arXiv pageView PDFProject pageAdd to collection

Get this paper in your agent:

hf papers read 2609\.03796

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper4

#### inclusionAI/LLaDA-Image-Turbo Text-to-Image• 7B• Updatedabout 2 hours ago • 2 • 3 #### inclusionAI/LLaDA-Image Text-to-Image• 7B• Updatedabout 2 hours ago • 5 • 2 #### inclusionAI/LLaDA-Image-Turbo-FP8 Text-to-Image• 7B• Updatedabout 2 hours ago • 2 #### inclusionAI/LLaDA-Image-FP8 Text-to-Image• 7B• Updatedabout 2 hours ago • 1

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.03796 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.03796 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles

LLaDA2.X (GitHub Repo)

TLDR AI

Ant Group releases LLaDA2.X series of diffusion language models, scaling to 100B parameters with MoE architecture, open-sourcing weights and training code.

DALL·E: Creating images from text

OpenAI Blog

OpenAI introduces DALL·E, a 12-billion parameter transformer model that generates images from text descriptions by treating text and images as a single token stream. The model demonstrates diverse capabilities including creating anthropomorphized objects, combining disparate concepts, rendering text, and performing image inpainting tasks.

HiDream-ai/HiDream-O1-Image-Dev

Hugging Face Models Trending

HiDream-ai has released HiDream-O1-Image-Dev, an 8B parameter open-source image generation model that uses a pixel-level unified transformer without external VAEs. It ranks #8 in the Artificial Analysis Text to Image Arena and supports high-resolution generation up to 2,048x2,048.

New BEST local AI image generator is here!

YouTube AI Channels

Ernie Image, a new open-source diffusion model, surpasses Zage in text rendering and prompt fidelity and can be run locally via ComfyUI with ~20 GB VRAM.