Sampler LLaDA-Image Turbo

Here's the thing about LLaDA-Image Turbo: it's a distilled model. It isn't a smaller, faster Base - it's a version where the guidance and the multi-step trajectory were trained into the weights, so it renders properly in 4 diffusion steps at CFG 1. That short trajectory needs its own sampler, one that re-injects a little noise at each step instead of just marching deterministically. That sampler is what this node hands you.

It has no inputs and no settings. Just one output, SAMPLER, which you plug into SamplerCustomAdvanced. It's the least glamorous node in the pack and also the one you must not quietly delete, because the moment you do, your Turbo workflow falls back to a sampler the checkpoint wasn't built for and the results go off.

How it works

The Turbo checkpoint's official path uses a stochastic flow sampler - per step it denoises, then mixes the cleaned signal with fresh noise before the next step (x = (1 − σ_next)·denoised + σ_next·noise, roughly). That stochasticity is what lets a 4-step flow model avoid the banding and flatness you'd expect from so few steps. The node's own description spells out the alternative plainly: pair the LLaDA-Image schedule with the standard Euler sampler for the official stochastic_sampling=false path. So you have two legitimate routes:

• Deterministic: KSamplerSelect → Euler, plus T8LLaDAImageScheduler for sigmas. Reproducible, and it's what Base uses anyway.
• Stochastic: this node's SAMPLER output, same scheduler, plus a seeded RandomNoise.

Both go into SamplerCustomAdvanced. The noise comes from the RandomNoise node, so changing the seed does what you expect.

What to actually set

The sampler is the one part you don't configure. The numbers around it matter more:

• CFG 1. Guidance is already baked into Turbo's weights. Crank CFG to 7 and you're applying the correction a second time - washed-out, blown images at twice the render cost. This is the same lesson Z-Image Turbo and Klein taught: on a guidance-distilled model, "CFG 1" is the correct setting, not a timid one.
• Steps 0 on the scheduler, which resolves to 4.
• Use it with the Turbo checkpoint. Feed it a Base model and you're running Base's long trajectory through a sampler tuned for 4 steps. Wrong tool, wrong results.

Install

Part of the LLaDA-Image T8 pack - nothing separate to fetch. Manager search "LLaDA-Image T8", or git clone https://github.com/T8mars/Comfyui-LLaDa-Image-T8 into custom_nodes/, pip install -r requirements.txt, restart. ComfyUI 0.34.0+ required.

The gotcha nobody reads twice

"Turbo's 4 steps" refers only to the diffusion stage. If you're doing VQ or image-editing workflows, the autoregressive LLaDA2 stage that runs before diffusion is still slow - the whole job will not take four seconds. That's not a broken install, it's the architecture. The sampler saves you time on the diffusion phase and leaves the rest of the pipeline alone.

Also worth knowing: because this sampler is stochastic by design, two runs with the same seed are still your friend for comparing prompts - but if you need pixel-exact reproducibility for something like a batch test, use the Euler path instead and keep your sanity.