LLaDA-Image

Welcome to the official repository for LLaDA-Image, a unified model for high-quality image generation and editing.

Introduction

LLaDA-Image is a competitive 6B-parameter open-source unified image generation and editing model family. It includes LLaDA-Image, a 50-step Base model for high-quality text-to-image generation and instruction-guided editing, and LLaDA-Image-Turbo, a 4-step distilled model for fast generation and editing. Both variants support practical text-to-image generation, VQ-conditioned generation, reference-image editing, and Chinese--English text rendering.

This repository provides the checkpoints and Diffusers-based inference code for the LLaDA-Image model family.

News

• 2026-09-04: We released the LLaDA-Image Base and Turbo checkpoints together with the inference code.

Highlights

• Unified generation and editing. A single checkpoint supports text-to-image generation and reference-preserving, instruction-guided editing without a separate editing backbone.
• Unified diffusion model. Both backone and DiT are diffusion models, trained in a unified framework.
• Realistic image generation. LLaDA-Image produces high-quality images with rich visual details, natural lighting, and coherent compositions.
• Image-only pre-training for visual-prior learning. The report establishes the visual prior through image-only pre-training and mid-training before introducing paired language supervision and joint generation--editing training.
• Efficient inference with distilled model. LLaDA-Image-Turbo uses Twin-DMD distillation to deliver fast image generation and editing in only 2--4 sampling steps.
• SOTA on Qwen-Image-Bench. LLaDA-Image achieves state-of-the-art overall scores of 53.53 in English and 53.38 in Chinese.

Model Zoo

Opensource Plan

• [x] Inference code and model weights
• [ ] Training code (coming soon)

Quick Start

1. Create an environment

The implementation has been used with Python 3.11, PyTorch 2.8, Transformers 4.57.6, and Diffusers 0.39.0.

git clone https://github.com/inclusionAI/LLaDA-Image.git
cd LLaDA-Image

conda create -n llada-image python=3.11 -y
conda activate llada-image

pip install -r requirements.txt

2. Run inference

The pipeline accepts a prompt and, for editing, an optional reference image.

LLaDA-Image (Base)

Use the Base checkpoint for high-fidelity generation and editing. Its recommended sampling configuration is 50 steps.

import torch

from src import LLaDAImagePipeline

# Load the pipeline. The model is downloaded from Hugging Face on first use.
pipe = LLaDAImagePipeline.from_pretrained(
    "inclusionAI/LLaDA-Image",
    torch_dtype=torch.bfloat16,
    device="cuda",
)

# Generate an image.
prompt = (
    "A cinematic photograph of a red fox standing in fresh snow, "
    "soft winter light, detailed fur, shallow depth of field"
)
negative_prompt = ""

image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    generation_mode="text",
    height=1024,
    width=1024,
    num_inference_steps=50,
    guidance_scale=5.0,
    generator=torch.Generator("cuda").manual_seed(42),
).images[0]

image.save("llada-image-base.png")

LLaDA-Image-Turbo

Use the Turbo checkpoint for fast generation and editing. Its recommended sampling configuration is 4 steps.

[!NOTE] For LLaDA-Image-Turbo inference, you can try setting stochastic_sampling to false in scheduler/scheduler_config.json, which may produce sharper details in some cases.

import torch

from src import LLaDAImagePipeline

# Load the distilled Turbo checkpoint.
pipe = LLaDAImagePipeline.from_pretrained(
    "inclusionAI/LLaDA-Image-Turbo",
    torch_dtype=torch.bfloat16,
    device="cuda",
)

prompt = "A quiet observatory above a sea of clouds at sunrise, golden light, wide-angle photograph"

image = pipe(
    prompt=prompt,
    generation_mode="text",
    height=1024,
    width=1024,
    num_inference_steps=4,
    guidance_scale=1.0,
    generator=torch.Generator("cuda").manual_seed(42),
).images[0]

image.save("llada-image-turbo.png")

Generation modes

Both checkpoints support the following modes. Text and VQ-conditioned generation require height and width divisible by 16; image editin