Ant Group (inclusionAI) · 2026-09-04 · major
LLaDA-Image — a 6B open image generator with a 4-step turbo variant
LLaDA-Image is a 6B open-weight model from Ant Group's inclusionAI that both generates and edits images. The repository reports 53.53 in English and 53.38 in Chinese on Qwen-Image-Bench, and a distilled Turbo checkpoint runs in 4 steps.
A 6B diffusion model from Ant Group that writes images from text, edits them from a reference, and publishes its training recipe.
Key specs
| Parameters | 6B |
|---|
Quick facts
| Maker | Ant Group (inclusionAI) |
|---|---|
| Released | September 4, 2026 |
| Checkpoints | LLaDA-Image (50 steps) and LLaDA-Image-Turbo (4 steps), each in BF16 and FP8 |
| Tasks | Text-to-image, VQ-conditioned generation, reference-image editing, bilingual text rendering |
| Qwen-Image-Bench | 53.53 English / 53.38 Chinese |
| Training scale | 220M samples, 98M of them real images |
| Out now | Weights and inference code; training code listed as coming soon |
What is it?
LLaDA-Image brings image generation and editing to the LLaDA diffusion family that Ant Group's inclusionAI team develops. One 6B open-weight model covers text-to-image prompts, VQ-conditioned generation, edits driven by a reference image, and text rendering in both Chinese and English. Two checkpoints went up on September 4: a 50-step base model and LLaDA-Image-Turbo, a distilled version that runs in 4 steps.
How does it work?
A 6B diffusion transformer does the generating, paired with a frozen vision-language module that reads the prompt. Training ran over 220 million samples, 98 million of them real images, using parameter-free RMSNorm and the Muon optimizer. LLaDA-Image-Turbo is distilled from the base model, which is how it cuts sampling from 50 steps down to 4.
Why does it matter?
Most open image models ship weights and stop there. The LLaDA-Image paper says the team releases the model weights, the training code and the detailed recipes, so a lab can retrain or adapt the model instead of only fine-tuning a black box. The repository reports state-of-the-art scores among open-source models on Qwen-Image-Bench, and FP8 copies of both checkpoints lower the memory needed to load them.
Who is it for?
teams building image generation and editing
Frequently asked questions
- What is the difference between LLaDA-Image and LLaDA-Image-Turbo?
- LLaDA-Image is the base checkpoint and runs 50 sampling steps. LLaDA-Image-Turbo is distilled from that base model and runs 4 steps, so it returns a picture much sooner while giving up the extra refinement passes. Both are published in BF16 and FP8 precision, and the FP8 copies need less memory to load.
- Can LLaDA-Image edit an existing picture, or only make new ones?
- LLaDA-Image does both. Alongside text-to-image prompts, the model supports instruction-guided editing driven by a reference image, as well as VQ-conditioned generation. The repository also lists bilingual text rendering, so LLaDA-Image can place readable Chinese and English characters inside the image it produces.
- Is the LLaDA-Image training code available yet?
- Not on day one. The September 4 release covers the LLaDA-Image and LLaDA-Image-Turbo checkpoints plus inference code, and the repository lists training code as coming soon. The accompanying paper states the team releases model weights, training code and detailed recipes to support further research on open generative models.
Try it
https://huggingface.co/inclusionAI/LLaDA-Image-Turbo