Dazzle TauSampler (DazzleNodes)

Dazzle TauSampler is a preview node, and the README is admirably direct about it: it implements a simplified form of tau complement sampling based on the Tau Operator from D. Darcy's Scarcity Framework. Concretely, at each denoising step it computes what the step just "removed" - the residual between the previous latent and the freshly denoised estimate, x_0 - x_next - and adds a controlled amount of it back. The idea is to recover information a standard denoiser discards, rather than letting that structure vanish forever. It's the most experimental thing this pack ships, and it's clearly labeled as such.

That "add a bit of what the step removed" framing is honestly easier to use than to fully understand. The sampler list is small and names itself: tau/res_2m, tau/res_2s, tau/dpmpp_2m, tau/dpmpp_2m_sde, tau/dpmpp_2s, tau/dpmpp_3m. The tau/ prefix means "this solver with the complement step bolted on" - the base solver underneath is the standard RES4LYF method, which is useful because it means you already know what the non-tau version looks like.

The inputs that matter

Three widgets define the behavior:

• tau_version - four flavors: tau1 (raw complement), tau2 (structural filtering, alignment + temporal - the default), tau3 (seed-variant perturbation), and tau4 (spectral per-bin complement via DFT - the newest and most experimental of the four).
• tau_strength - 0 to 1, default 0.5, and here's the reassuring part: internally it maps to an effective range of 0–0.12, and 0 is bit-identical to standard sampling. The node literally cannot change your output if you set it to zero.
• tau_mode - hard (fixed strength every step), soft (more complement in the low-noise detail phase, default), or cosine (a smooth ramp).

The rest of the widget set mirrors the flagship DazzleKSampler - sampler_name, scheduler (default beta57), steps, steps_to_run, denoise, cfg, seed, sampler_mode, bongmath, and latent_role with the same five-way noise-vs-init dispatch. Outputs are the standard triple: output (LATENT), denoised (LATENT), options (OPTIONS).

Installing it

It's part of the same pack:

cd ComfyUI/custom_nodes
git clone https://github.com/DazzleNodes/ComfyUI-DazzleKSampler.git

Restart ComfyUI and it's under RES4LYF/samplers, or install via ComfyUI Manager ("Dazzle KSampler"). No extra dependencies, no model files.

Where people get burned

The changelog's own caveat is worth taking seriously: the v1 implementation is a simplified complement (x_0 - x_next), and future versions plan proper structure/noise separation via a "resolution function R." Which is a long way of saying this is alpha research plumbing, not a polished effect. Expect subtle changes at low strengths and genuinely weird output if you crank it - and don't build a production workflow on it. The right way to play with it: keep tau_strength low, A/B against the plain DazzleKSampler on the same seed and settings, and treat anything above maybe 0.3 as "interesting, not necessarily better." The safe lane here is real: tau_strength = 0 is a perfect drop-in that behaves exactly like the standard sampler, so you can experiment without risk of corrupting a workflow you care about.