Apppublic

hugging-apps/llada-image-turbo-demo

sourceHugging Faceupdated 23d agoView on Hugging Face

17likes

App README

LLaDA-Image-Turbo

Demo of **LLaDA-Image** — Building Strong Image Generators with Fully Open Training Recipes (inclusionAI) — in its Turbo configuration, distilled with Twin-DMD down to 4 sampling steps.

The model is a 6B flow-matching DiT conditioned by a frozen LLaDA2 MoE understanding backbone, and handles both text-to-image generation and instruction-guided image editing in a single set of weights. Prompts work in English and Chinese, including rendered text.

—Paper: https://huggingface.co/papers/2609.03796
—Code: https://github.com/inclusionAI/LLaDA-Image
—Weights: `inclusionAI/LLaDA-Image-Turbo` and `inclusionAI/LLaDA-Image-Turbo-FP8`

How this Space loads the model

The full bf16 Turbo checkpoint is ~49 GB, which does not fit in a single ZeroGPU slot, so the two official Turbo releases are combined:

Component	Source
`text_encoder` (LLaDA2 MoE)	`inclusionAI/LLaDA-Image-Turbo-FP8` — the authors' blockwise-FP8 weights (~17 GB)
`transformer`, `vae`, `sigvq`, `queryformer`, `text_projection`	`inclusionAI/LLaDA-Image-Turbo` (bf16)

The FP8 transformer in the -FP8 repo uses the fused SGLang module layout (to_qkv / w13), which the authors' released diffusers reference code cannot load, so the bf16 DiT is used. Everything else follows the official pipeline in src/ (vendored unchanged from the GitHub repo).

Sampling defaults follow the model card's Turbo recipe: 4 steps, guidance scale 1.0.

Example images

Example inputs for the editing tab come from `linoyts/repo-to-space-example-inputs` (CC0-1.0).