LLaDA-Image merges image generation and editing in a 6B open model
inclusionAI just open-sourced LLaDA-Image, a 6B unified diffusion model family that handles both text-to-image and instruction-guided editing from a single checkpoint. It delivers photorealistic rendering alongside bilingual Chinese-English text, while a 4-step distilled Turbo version enables near-instant generation. Checkpoints and ComfyUI integrations are already live.
LLaDA-Image packs high-quality text-to-image generation and instruction-guided editing into a single 6-billion-parameter open-source model.
Usually, generating an image from scratch and modifying an existing one require separate systems or heavy auxiliary networks. inclusionAI built a unified diffusion framework where a single checkpoint handles pristine scene generation, bilingual text rendering, and reference-based visual tweaks without extra editing backbones.
What it does
LLaDA-Image covers both end-to-end rendering and precision editing.
On the generation side, it focuses on photorealistic lighting, cohesive composition, and sharp Chinese-English text rendering, making it practical for posters and information graphics.
Its main hook is instruction-based editing. Feed it a source picture with a simple prompt, and it preserves unchanged areas while tweaking specific elements. You can modify character expressions, switch postures, swap objects, rewrite in-image text, or colorize black-and-white photos.
On Qwen-Image-Bench, the base model achieved overall scores of 53.53 in English and 53.38 in Chinese.
The release includes two model variants: the standard LLaDA-Image Base running at 50 sampling steps for high-fidelity outputs, and LLaDA-Image-Turbo, which uses Twin-DMD distillation to produce results in just 2 to 4 steps.
How to run it
Getting it running locally is straightforward and relatively lightweight on hardware.
Checkpoints are available on Hugging Face and ModelScope in both BF16 and FP8 formats. The FP8 version of the Turbo Transformer measures roughly 6.7GB, meaning it can easily squeeze into typical consumer GPUs.
The codebase relies on Python 3.11, PyTorch 2.8, and Diffusers 0.39.0. Text-to-image and VQ-conditioned modes require image dimensions divisible by 16, while reference-image editing requires dimensions divisible by 32.
For those who prefer visual node setups over Python scripts, community ComfyUI integration landed on September 7, fully supporting FP8 workflows. Model weights and Diffusers inference scripts are fully accessible now, with training recipes scheduled to follow.
Why it matters
For creators and local AI hobbyists, LLaDA-Image eliminates the headache of swapping heavy checkpoints back and forth just to tweak a generated image. Getting competent text rendering, realistic generation, and instruction editing in a sub-7GB FP8 footprint makes it ideal for local desktops.
By collapsing generation and editing into a unified DiT diffusion model, it sets a practical benchmark for compact multimodal tools. While full training recipes are still pending release, the 4-step Turbo checkpoint alone provides an efficient everyday asset for creators running on constrained VRAM.
Sources (1)
- github.com(opens in a new tab)Originalgithub.com