ModelSamplingAuraFlow

ModelSamplingAuraFlow is a single shift slider that most people meet because a shared workflow told them to set it to 7. And for once the shared workflow is right: on Z-Image it's genuinely the difference between flat, washed-out output and something with real contrast and life. It's the most-searched node in this whole model/patch family for a reason - it's the one patch that the Z-Image community treats as basically mandatory.

The name comes from its original target, AuraFlow, the 6.8B rectified-flow transformer fal shipped in July 2024 under Apache 2.0. That model is a footnote now, but the patch survived because it does something useful for a whole family of flow-matching models. Don't over-think the name - think of it as "retune the noise schedule for a flow model."

How it works

Diffusion models denoise in stages, and which stages get the most sampling steps matters enormously. A flow-matching model like Z-Image or AuraFlow predicts a straight-ish trajectory from noise to data, and the shift parameter bends the mapping between "how much noise is left" and "which timestep you're on." Raise the shift and the sampler spends more of its steps in the high-noise region where broad composition gets decided; lower it and more steps go to the fine-detail end.

Under the hood it clones your model and swaps in a ModelSamplingDiscreteFlow sampler with a CONST sampling type, then calls set_parameters(shift=...) - the same machinery ModelSamplingSD3 uses, just with a multiplier of 1.0 instead of 1000 because the timestep scale is different. Nothing else in your graph needs to change.

Inputs and what to set

Just two inputs, both on the MODEL it hands to a KSampler:

• model - wire it between your checkpoint/UNet loader and the sampler.
• shift - float, default 1.73. This is the one you touch. On Z-Image Base and Turbo, the community's consistent finding is that 7 gives noticeably better results than the default - it was one of the most-recommended tweaks across the early Z-Image threads, and it's why every Z-Image template workflow you'll find has this node set to 7.

The single output is the patched MODEL, which feeds your sampler like the unpatched one would. There's no model file to download; this node ships with ComfyUI and does its work purely in math.

Common issues & troubleshooting

The number does different things on different models. A shift that's right for Z-Image can wash out or overcook another flow model. If you grabbed a workflow for a different architecture, don't trust the 7 - AuraFlow itself wants something near the 1.73 default, and SD3-family models use this node's sibling at 3.

It does nothing visible on a distilled model at CFG 1. Because the schedule is the one thing a distilled, few-step model has already optimized, the effect shrinks as steps drop. If you're on a Turbo/8-step path and see no difference, that's expected - the shift matters most at 20+ steps.

Order matters if you stack patches. Patch nodes chain in sequence; if you're also applying a LoRA or a CFG patch, put the model-sampling patch closest to the loader or sampler whichever your workflow author intended. The patch only affects the model object it's applied to, so re-arranging the chain genuinely changes results.