How to Run LLaDA-Image Locally

01

1. Choose Base or Turbo

Start with LLaDA-Image Base when fidelity matters more than inference time. Start with LLaDA-Image-Turbo when you need faster iteration. The official examples use 50 steps for Base and 4 steps for Turbo.

Both variants have BF16 and FP8 checkpoint repositories. FP8 can be relevant for lower-precision deployments on compatible hardware, but the existence of FP8 weights does not establish a guaranteed VRAM requirement.

02

2. Reproduce the official environment

The repository documents Python 3.11, PyTorch 2.8, Transformers 4.57.6 and Diffusers 0.39.0 as its used stack. Clone inclusionAI/LLaDA-Image, create a Python 3.11 environment and install the repository requirements before debugging with newer dependency combinations.

The released examples run on CUDA. That confirms the documented execution path, but it is not a universal statement about one required GPU model or memory capacity.

03

3. Run Base or Turbo text generation

Load the desired checkpoint with LLaDAImagePipeline.from_pretrained. Base uses the inclusionAI/LLaDA-Image checkpoint with the repository’s 50-step starting configuration; Turbo uses inclusionAI/LLaDA-Image-Turbo with the 4-step starting configuration.

04

4. Use VQ generation or image editing

For VQ-conditioned generation, set generation_mode='vq' and do not pass an input image. The official docs say LLaDA2 produces image VQ tokens from the prompt and SigVQ embeds them before diffusion. VQ dimensions must be divisible by 16.

For instruction-guided editing, load a reference image and use generation_mode='editing'. Editing dimensions must be divisible by 32. The same checkpoint family supports both creation and reference-image edits.

05

5. Measure hardware instead of guessing

inclusionAI does not currently publish one minimum VRAM figure covering every model, precision, resolution and mode. Measure peak memory on the actual GPU and settings you plan to use, then test Turbo or FP8 if speed or capacity is limiting.

• Record model variant and precision.
• Record resolution and generation mode.
• Record sampling steps and guidance scale.
• Measure runtime and peak memory.
• Treat third-party VRAM numbers as hardware-specific evidence, not official requirements.

06

6. Avoid common setup mistakes

• Do not use arbitrary dimension sizes: text/VQ require multiples of 16 and editing requires multiples of 32.
• Do not run Turbo with Base-style 50-step settings when the goal is fast distilled inference.
• Do not pass an input image in VQ mode.
• Do not assume public weights automatically grant commercial-use rights.
• Do not describe training code as released while the repository still marks it coming soon.

For creators, a sensible workflow is to validate Turbo first for prompt iteration, test editing on representative assets, then compare Base only where the quality difference matters enough to justify the extra inference cost.