Latent Disintegration (S42)

S42_LatentDisintegration is the "break it on purpose" node. It takes any LATENT tensor and injects noise directly into the semantic space, producing a controlled decay - the visual equivalent of a signal dissolving into static, but happening in the latent before the VAE decodes, not in the pixels after. If you've ever wanted footage that slowly falls apart into abstract mush, this is the knob.

It lives in the "Experimental" corner of the pack's audio/latent section, and here's the thing that makes it broadly useful: it doesn't care what kind of latent you feed it. Image latents, video latents, LTX-2 audio latents - they're all just tensors to this node, and the disintegration works on any of them.

How it works

The mechanism is embarrassingly simple, and that's a feature: samples = samples * (1 - entropy_level) + noise * entropy_level. At entropy_level 0, the latent passes through untouched. At 0.2 (default), it's 80% original, 20% random noise - a subtle structural fuzz. At 0.5, the signal and the noise are fighting. At 1.0, it's pure noise, which decodes to... garbage. But the garbage is the point - it's how you get a "dissolving into nothing" moment.

Because it operates on the latent rather than pixels, the artifacts look organic - they smear through the model's learned structure rather than sitting on top of the frame like a filter.

Inputs

• latent - any LATENT. In an LTX workflow, grab it from EmptyLTXVLatentVideo or an encoder and stick this node in the path before the sampler/decoder.
• entropy_level - 0–1, default 0.2.

That's the whole API. One input, one knob, one output. Refreshing.

Output

A single LATENT, ready to feed a KSampler or straight to a VAE decode. Wire it into the normal latent pipeline and the disintegration carries through to the frames.

Installing it

Standard S42 CutFlow install: ComfyUI Manager (search "S42 CutFlow") or git clone https://github.com/GeekyGhost/S42-CutFlow into ComfyUI/custom_nodes/, install requirements, restart. Pure torch tensor math - no models, no downloads, runs anywhere.

Gotchas

The honest caveats: there's no seed input, so the noise comes from torch's global RNG state. Two runs back-to-back will differ unless you've seeded the workflow; if you need it reproducible, fix a seed upstream. And watch the level - the step between 0.4 and 0.6 is where it goes from "distressed footage" to "snow." For a slow disintegration over a clip, feed it a scheduled entropy_level from a curve node rather than a fixed value, and pair the result with the pack's LatentDisintegration-adjacent experimental nodes (or just decode it and A/B) to see what's actually happening before you commit.