OCS Group

The whole point of the Overly Complicated Sampling pack is doing more than one thing during a single sampling pass, and OCS Group is the node that makes that possible. It's the "when" switch in the OCS stack: it decides which set of substeps runs at which step of the denoise, so you can run euler for the first few steps, switch to dpmpp_sde in the middle, and finish on something stable. Stock ComfyUI can't do that without hacky custom_sampling patches; OCS gives it to you as a chain of ordinary nodes.

What it is

A OCS Group sits between the sampler and the substeps: OCS Sampler ← OCS Group ← OCS Substeps. The group defines a window of time during sampling and a way of merging its substeps. Groups chain together through the optional groups_opt input, and matching starts with the group furthest from the sampler node - the first one in the chain wins for a given step.

The inputs that matter

You'll actually set three of these:

• merge_method - how multiple substeps combine into one step. simple (default) runs a single substep per step. divide splits the step into a linear schedule of substeps, one model call each. supreme_avg runs the substeps sharing the first model call and averages the results. overshoot samples ahead and then restarts back to the expected step - this is the one most people try first, and it's where OCS's built-in restart support comes in. lookahead and pingpong are experimental variants of the same idea. dynamic lets an expression pick the method per step.
• time_mode + time_start / time_end - step (default), step_pct, or sigma. Matches are inclusive, and steps start at 0. step mode is by far the easiest to reason about.
• substeps (required) - connect the OCS Substeps node here.

Outputs are OCS_GROUPS, which you feed either to another Group's groups_opt or straight into the OCS Sampler. Optional params_opt and the parameters text block work like everywhere else in OCS - the text block is where pre_filter and post_filter live, along with preview_mode and the when expression for time-based gating beyond the simple range.

A concrete example

Want euler for the first three steps and dpmpp_2m afterward? Make two groups. The first has time_start: 0, time_end: 3, and a Substeps node set to euler. The second has time_start: 4, time_end: 999, and a Substeps node set to dpmpp_2m. Chain Group1 -> Group2 -> Sampler, and the sampler picks per step. That's the entire skill: groups are just time-gated wrappers.

Installing and gotchas

Installed as part of the pack - ComfyUI Manager → search "Overly Complicated Sampling", or git clone https://github.com/blepping/comfyui_overly_complicated_sampling into ComfyUI/custom_nodes and restart. No models, no mandatory pip deps.

Two things bite people. First, supreme_avg doesn't work well with non-ancestral samplers (anything with eta: 0), and it's pointless to run identical substeps since they share the first model call anyway. Second, the pack's restart-based merge methods (overshoot, lookahead) don't work with rectified-flow models like Flux. And remember the pack is "in flux" - merge methods are exactly the kind of thing that changes between versions, so re-check saved workflows after updates.