DiT Spectrum Patch

If you already have a working sampling graph - stock KSampler, KSampler Advanced, a Custom Sampler chain with other wrappers - and you don't want to rebuild it around a new sampler node, this is the way in. DiT Spectrum Patch is Spectrum's core caching trick exposed as a plain MODEL → MODEL patcher: wire it before whatever sampler you already use and it accelerates that sampler transparently.

CheckpointLoader / model loader → DiT Spectrum Patch → KSampler / KSampler Advanced / Custom Sampler

What it does, and - importantly - what it doesn't

This node is the caching mechanism only, stripped down to nothing else. Same family as TeaCache-style step-caching accelerators: it forecasts DiT feature outputs instead of computing every step in full. Specifically, on cached steps it predicts the transformer feature immediately before Anima's final_layer from a Chebyshev ridge-regression fit over recent actual steps, then runs only final_layer + unpatchify - skipping every transformer block for that step. On actual steps, the full model runs and its feature gets captured to feed the forecaster.

What it explicitly does not do: no modulation guidance, no DCW, no SMC-CFG, no SPEED/SPD multi-resolution, no noise generation or latent padding, no custom sampling loop. It's a caching wrapper and nothing more - for the correction stack, chain DiT CFG-FSG/DCW Patch or Anima Mod Guidance separately.

Inputs and outputs that matter

• steps - has to match your downstream sampler's own step count exactly. This is the one setting most likely to bite you: change your sampler's steps and forget to update this node, and the forecaster's internal schedule desyncs from what the sampler is actually doing.
• window_size / flex_window / warmup_steps - the same adaptive caching schedule as the integrated samplers: warmup_steps forces full compute up front, then an actual forward runs every floor(window_size) cached steps, growing by flex_window after each real forward. Raise flex_window for more speed, lower it (or raise warmup_steps) for more fidelity.
• tail_actual_steps (default 3) - the final N steps are always forced to actual forwards, protecting the last stretch of denoising (where detail resolves) from ever being a forecasted guess.
• enabled - flip to false to pass the input model through completely unchanged, a clean kill switch for A/B testing.
• one_sampler_only - if the same patched model feeds multiple samplers in one workflow run (a hi-res-fix second pass, for instance), setting this true applies Spectrum only to the first sampler that consumes it; later ones in that same run pass through untouched. It re-arms automatically on the next full workflow execution.
• verbose - logs the actual/cached decision per step, useful for confirming the schedule is doing what you expect.

Output is a patched MODEL.

Installing it

ComfyUI Manager - search SpectrumKSampler, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/sorryhyun/ComfyUI-Spectrum-KSampler

Restart - no model weight downloads for this node specifically; it's pure computation, no auto-fetched assets.

Common issues & troubleshooting

Output looks subtly off, or the schedule seems desynced. Check steps first - it has to match your sampler's step count. This is the single most common way to misconfigure this node.

Spectrum applied to a hi-res-fix pass when you didn't want it to. If the same patched model feeds two sampler nodes in sequence, both get accelerated by default. Set one_sampler_only = true to limit it to the first run.

Node errors instead of producing garbage on a non-DiT model. That's deliberate - the pack fails loudly with a clear error on non-DiT models rather than silently producing invalid output, so if you see that error, it means you've wired this into a checkpoint it genuinely can't accelerate, not that something else went wrong.

Need it to play nicely with a chained attention-backend wrapper or Custom Sampler flow. This basic node has no compatibility controls - for that, use DiT Spectrum Patch Advanced, which adds a compat_policy setting specifically for wrapper-heavy or multi-conditioning graphs.