OCS Sampler

If you install this pack for one node, this is it. OCS Sampler is the flagship: a drop-in replacement for the sampler you normally pick in the KSampler dropdown, except instead of one algorithm it can run a different algorithm every few steps, restart the schedule, use immiscible noise, and filter almost anything mid-sampling. The README calls the pack "experimental and mathematically unsound (but fun!)" - read that as the tone for everything below. This is not the node you reach for when you want a quiet reliable euler run; it's the node you reach for when you want to know how far sampling can be bent.

What it is

OCS doesn't plug into a KSampler. It emits a SAMPLER that you wire into a SamplerCustom (or anything that accepts a SAMPLER input). Structurally it's the top of a three-node stack: OCS Sampler ← OCS Group ← OCS Substeps. The Group nodes decide when something runs, the Substeps nodes decide what sampler runs, and the Sampler node holds the global config - noise handling, restart, model options - plus everything the groups couldn't express.

How it works

The sampler node turns your chain of groups into a single custom KSAMPLER. Every step, it asks the group chain "who matches this step?" and runs that group's substeps. Matching is time-based and starts with the group furthest from the sampler node, so later groups in the chain are tried first. The heavy lifting lives in a YAML/JSON text block where you can set the noise block (cpu_noise, set_seed, seed_offset, normalize_noise), the restart block (enabled plus its own s_noise and immiscible config), and the model block - the one beginners actually need.

The model block's cfg1_uncond_optimization is the Flux switch. Flux doesn't use CFG, so setting this to true skips generating the uncond (negative prompt) pass entirely, saving time. The flip side: CFG++ and alt CFG++ features don't work with Flux, and restarts don't work with rectified-flow models like Flux at all.

The inputs that matter

The node is refreshingly simple at the wire level:

• groups (required) - connect your OCS Group chain here.
• params_opt (optional) - output from OCS Param / OCS MultiParam for things that need a real input, like custom noise.
• parameters (optional) - the big YAML text block where all the tuning happens.

The single output, SAMPLER, goes into SamplerCustom. From there it behaves like any sampler: add_noise, start_at_step, and so on are handled by the SamplerCustom node.

Installing it

From ComfyUI Manager, search for Overly Complicated Sampling. Or clone it manually:

cd ComfyUI/custom_nodes
git clone https://github.com/blepping/comfyui_overly_complicated_sampling

Restart ComfyUI. There are no model files to download and no mandatory pip dependencies - the optional solver backends (diffrax, torchdiffeq, torchode, torchsde) are only imported if you pick those step methods. The author strongly recommends installing ComfyUI-bleh and ComfyUI-sonar alongside, since they unlock extra blend modes and noise types.

Common issues

The pack's own status is "in flux, may be useful but likely to change/break workflows frequently." Treat saved workflows from old versions as fragile - an update can silently change a default. The YAML block has "essentially no error handling," so a typo surfaces as a confusing traceback rather than a friendly message. And if you're on Flux, remember the cfg1_uncond_optimization note above before you start diagnosing weird slowness. This is a power-user tool that rewards you only if you're willing to read the README's YAML reference; if you just want DPM++ 2M Karras with fewer steps, the built-in samplers are still your friend.