Ant's inclusionAI Open-Sources LLaDA-Image, a 6B Unified Image Generation & Editing Model
On September 4, Ant Group's inclusionAI open-sourced LLaDA-Image, a 6B-parameter unified image generation and editing model. The paper is on arXiv, weights are hosted on Hugging Face, and code is on GitHub, alongside a distilled fast variant LLaDA-Image-Turbo. The team claims open-source SOTA performance, with a training recipe notable for not relying on large-scale image-text pairs.
Confirmed
- LLaDA-Image is a 6B Diffusion Transformer trained from scratch, unifying text-to-image generation and instruction-guided editing within a single framework, with both the backbone and image generation based on diffusion models.
- It pairs with a frozen vision-language understanding module built on the LLaDA2.0-Mini diffusion language model backbone; both the text encoder and the backbone are diffusion models.
- Training uses pure-image pretraining and the Muon optimizer; per @inclusionAI and @iScienceLuvr, it does not depend on large-scale image-text paired data.
- A 2-4 step distilled fast variant, LLaDA-Image-Turbo (listed on Hugging Face as a 4-step distillation), is provided and claimed to be competitive at this scale.
- The team states it can produce photorealistic images with precise editing, reaching open-source SOTA.
Why it matters
- By unifying generation and editing in one diffusion architecture with a fully open training recipe, it offers the community a reproducible technical path; using a diffusion language model as the text encoder also extends the boundaries of diffusion-model applications.
2026-09-04 ~ 2026-09-04 · 5 related posts
Primary sources
- LLaDA-Image: 6B diffusion model with Muon optimizer hits open-source SOTA image generation — inclusionAI ·
- LLaDA-Image: 6B diffusion model with fully open training recipes for photorealistic images — iScienceLuvr ·
- LLaDA-Image: unified 6B image generation/editing model released open-source — Total-Resort-3120 ·
- [source] LLaDA-Image: 6B diffusion model with Muon optimizer hits open-source SOTA image generation — inclusionAI · 2026-09-04
- InclusionAI releases LLaDA-Image: one 6B diffusion DiT for generation and editing — bdsqlsz · 2026-09-04
- [source] LLaDA-Image: unified 6B image generation/editing model released open-source — Total-Resort-3120 · 2026-09-04
- [source] LLaDA-Image: 6B diffusion model with fully open training recipes for photorealistic images — iScienceLuvr · 2026-09-04
- LLaDA Image Turbo 6B lands: 4-step distilled unified diffusion model, open weights — Gradio · 2026-09-04