---
title: LLaDA-Image
canonical_url: "https://www.modelscope.ai/studios/Iceberg-Yang/LLaDA-Image"
md_url: "https://www.modelscope.ai/studios/Iceberg-Yang/LLaDA-Image.md"
repository: Iceberg-Yang/LLaDA-Image
chinese_name: LLaDA-Image
last_updated: 2026-09-07
sdk_type: gradio
sdk_version: 6.17.3
downloads: 0
stars: 1
---

# LLaDA-Image

> LLaDA-Image - An interactive demo by Iceberg-Yang on ModelScope. English demo of inclusionAI/LLaDA-Image: unified text-to-image and instruction-based editing (Turbo 4-step / Base 50-step), running on ModelScope xGPU

Iceberg-Yang/LLaDA-Image is an interactive demo (Studio) on ModelScope. built with gradio 6.17.3.

- **Repository**: Iceberg-Yang/LLaDA-Image
- **SDK**: gradio
- **SDK version**: 6.17.3
- **Downloads**: 0
- **Stars**: 1
- **Last updated**: 2026-09-07

Source: https://www.modelscope.ai/studios/Iceberg-Yang/LLaDA-Image

---

# LLaDA-Image — ModelScope Studio (English)

English-language demo Space built on [inclusionAI/LLaDA-Image](https://www.modelscope.ai/models/inclusionAI/LLaDA-Image),
running on ModelScope xGPU (L20 48GB). Checkpoints are pulled from the international hub (modelscope.ai).

## Features

- **Text to image**: bilingual prompts, strong on in-image text rendering and photorealism
- **Image editing**: reference image plus an instruction, preserving subject identity
- **Two modes**: Turbo (4-step fast) / Base (50-step quality), single-choice switch

## Backend

- Inference code: the official [GitHub inclusionAI/LLaDA-Image](https://github.com/inclusionAI/LLaDA-Image)
  `src/` package @ `abd68bd`, vendored as-is with no changes to the inference logic
- Weights: ModelScope FP8 variants (Turbo-FP8 / Base-FP8). The DiT is block-dequantized to bf16 at
  load time, while the text encoder keeps the official FP8 expert path (`torch._scaled_mm`, with an
  automatic fallback to the reference implementation when unavailable)
- VRAM strategy: one variant stays resident on the GPU (~32 GiB), the other is parked on CPU and
  swapped in on demand; on OOM the text encoder automatically degrades to CPU execution
- Cache: `/mnt/workspace/modelscope-cache`, so restarts skip re-downloading (~28 GB per variant)
- Paper: [arXiv 2609.03796](https://arxiv.org/pdf/2609.03796)

## Environment variables (all optional)

| Variable | Default | Description |
|---|---|---|
| `LLADA_PRECISION` | `fp8` | `fp8` or `bf16` (bf16 keeps the text encoder on CPU) |
| `LLADA_MOE_BACKEND` | `auto` | MoE kernel: `auto` / `triton` / `eager` |
| `LLADA_KEEP_BOTH` | `0` | When `1`, inactive variants are parked in host RAM (second-level restore; needs a host with ≥96GB RAM) |
| `LLADA_PREFETCH_ALL` | `0` | When `1`, all variants are preloaded at startup |
| `LLADA_WORKSPACE` | `/mnt/workspace` | Root directory for persistent caches |
| `MODELSCOPE_ENDPOINT` | `https://www.modelscope.ai` | Hub this Space downloads checkpoints from |

## License and notice

The model weights and the official inference code are copyright the inclusionAI/LLaDA-Image team
(the repositories ship no standardized LICENSE; permission to use them in this public demo Space was
granted by the official team). The glue code of this Space is released under Apache-2.0; files under
`src/` retain their original copyright headers, and `text_encoder/fused_moe_ops.py` is derived from
ByteDance-Seed/VeOmni v0.1.0.post1 (Apache-2.0).