BlehCFGInitSampler

The idea here is simple even if the name isn't: instead of running your sampler from step zero and throwing away the noisy early result, jump straight ahead to a later step with a latent that already approximates where that step would have left things. Fewer steps computed, same rough destination. The node's own description name-checks the technique it's closest in spirit to - CFGZeroStar's zero-init approach - and builds a small family of ways to do the skip.

There are four modes. zero is the simplest and most literal: it behaves like CFGZeroStar's zero init, which in practice just skips the configured number of steps outright with no attempt to compensate for what was skipped. afs stands for "analytical first step" - a formula the pack's own Credits section attributes to a published method (arXiv 2210.05475) - and it scales down the initial noise so it better approximates a latent that's already been partially denoised the right amount, rather than starting from raw, unscaled noise at a step where the schedule expects something further along. afs_flow_hack is the same idea adapted for flow-matching architectures, per the tooltip's own wording that it "may work better for flow models" - the same family of models where a lot of other sampler-level advice (schedulers, CFG defaults) also diverges from the older SD 1.5/SDXL playbook. scale_down is described in the tooltip as simply "the simplest approach to scaling down the latent to match the skipped steps" - cheaper to reason about than AFS, if less precise.

Inputs: sampler (required - the SAMPLER being wrapped), steps (default 1, range 0-9999 - how many steps to skip before the wrapped sampler takes over; fractional values are technically allowed but the tooltip itself warns they "may not work well," and whole-step skips are cheaper anyway since they don't require an extra model call to interpolate), mode (one of the four above, zero by default).

Output is a SAMPLER, ready to drop into SamplerCustomAdvanced or wherever your graph expects one.

Worth being upfront about: this node isn't named or explained anywhere in the pack's README prose - the mode dropdown's own tooltip text, plus the AFS credit line tucked into the README's separate Credits section, are the only real documentation available. Treat it as an advanced, experimental node rather than something with settled best-practice guidance: worth trying if you're chasing small speed wins on a longer sampling chain (particularly one already built from BlehInsaneChainSampler or a similar multi-sampler setup, where skipping the first few steps of the first link in the chain is a natural place to save time), but there's no community writeup or troubleshooting thread to lean on if a given mode behaves unexpectedly on your specific model.

Install is the standard route for the pack: ComfyUI Manager, or git clone https://github.com/blepping/ComfyUI-bleh into custom_nodes followed by a restart. No extra dependencies. If you're not sure which mode to start with, zero is the simplest to reason about and the closest to how CFGZeroStar itself behaves - save afs and its flow-model variant for once you've confirmed the basic skip is doing what you expect.