🔥 FlowBuilder

🔥 FlowBuilder is the whole point of the komojini pack. It's the node that lets you run only the part of a workflow that produces a given output - so you can generate "step by step" instead of re-rolling the entire pipeline every time you touch one knob.

Here's the problem it solves. A video or multi-pass workflow is a long chain: load → preprocess → sample → decode → save. When you're iterating, you usually only changed the tail - a denoise value, a prompt, a second pass - yet ComfyUI reruns the whole graph. If you're sampling video frames, that's minutes of GPU time burning on stages that produced identical results. FlowBuilder gives you a manual "run from here to this output only" switch.

How it works. You place a FlowBuilder node right before the output node you care about - PreviewImage, SaveImage, VideoCombine, whatever. In the UI, the node gets a Queue button. Click it and the pack's JavaScript does the surgical part: it mutes every node that doesn't feed into your FlowBuilder's output, queues the prompt, waits for it to finish, then un-mutes everything. Net effect: only the upstream subgraph that leads to that output runs. The optional batch_size input exists for running that flow multiple times, though the plain FlowBuilder mostly cares about the single run - the "advanced" variants add the loop UI.

Inputs and output. A single value input of any type (*) - wire it to whatever the output node needs - and an optional batch_size (INT, default 1). One * output that just passes the value through. There's no compute here; FlowBuilder is a routing and control node. It's also set to ALWAYS execute mode internally so it doesn't get caught up in the very muting it performs.

Why it pairs with the pack. This is a workflow from the early-video era: you build one big graph that can produce several outputs (a preview, a save, a video), and instead of maintaining separate workflows you keep one graph and run the branch you're working on. The cacheable samplers in the same pack are the companion trick - cache the expensive sampling so re-running a branch doesn't resample what didn't change.

Install. Part of komojini-comfyui-nodes:

ComfyUI Manager → search "komojini-comfyui-nodes" → Install → restart

or cd ComfyUI/custom_nodes && git clone https://github.com/komojini/komojini-comfyui-nodes. No models, no heavy deps beyond the pack's pytube/opencv requirements.

Where people get burned. The mute/unmute dance is the fragile part. It's a Feb-2024 pack, and the JS that mutes nodes and restores their state leans on older ComfyUI internals - on a current build, the mute/restore cycle is the most likely thing to misbehave. It also only restores nodes it muted, so if you manually muted something beforehand, don't be surprised if it comes back enabled. And remember this runs a real prompt on your GPU queue - it's a convenience, not a dry-run. For maintained "run subgraph" tooling, look at how rgthree's grouping nodes and the modern execute-selection features handle the same problem; FlowBuilder is the 2024 take.