ModelSamplingAuraFlow (Ray)

ModelSamplingAuraFlow (Ray) is a one-knob node that changes how a flow-matching model distributes its denoising effort, wrapped so it works on the Raylight distributed pipeline. If you've ever seen a video come out with decent composition but mushy detail - or sharp detail but a composition that drifted - this is the knob that rebalances those.

The mechanism is the sigma shift. Flow-matching models like AuraFlow (and by extension Flux, LTX, and friends, which share the same sampling math) spend their steps across a noise schedule. The shift parameter redistributes where those steps land: a higher shift pushes more steps into the early, high-noise phase (locking in global composition), a lower shift concentrates them in the late phase (refining fine detail). The knowledge-base crowd has a tidy way of putting it: shift is roughly the flow-matching replacement for choosing a scheduler, and it's model-specific - different checkpoints genuinely want different values.

So the input list is short on purpose:

• shift - default 1.73 (AuraFlow's canonical value). Raise it toward 2–3 if composition is unstable; lower it if output looks soft or overworked. It's cheap to sweep, so do that instead of guessing.

That's it. One float, one ray_actors in, one ray_actors out. The output feeds back into the chain that goes to your Ray sampler, because these "ModelSampling" nodes all work by patching the sampling object on the model before sampling happens.

The trap that trips everyone

The shift won't do anything if your scheduler ignores it. The ComfyUI concept docs warn about this specifically: some schedulers - bong_tangent is the named one - apply their own sigma shift and ignore the value in the workflow. So if you set shift, see zero change, and conclude the node is broken, check your scheduler first. Swap to a standard scheduler like Euler or one of the standard flow schedules and the knob starts behaving.

Anything else?

It's a raylight port of ComfyUI's stock ModelSamplingAuraFlow, which means there's a non-Ray equivalent you can compare against if you're troubleshooting. Install is the pack, no model downloads. And if you need the same idea but with a resolution-aware shift (the way Flux does max_shift/base_shift), the pack has Model Sampling LTXV (Ray) for the LTX flavor - this node is the fixed-shift variant.