Cherry Pick Frames

Every video pipeline eventually needs "just these frames" - the keyframes for a storyboard, the endpoints for a first/last-frame model, an even sample to preview a 400-frame render without watching all of it. ComfyUI's native batch slicing is clunky about this, which is exactly the gap Cherry Pick Frames fills: one node, four selection modes, and the outputs appear on the node face as you use them. It's plumbing, and it's the good kind.

The modes

• first_n - the first N frames. Default mode.
• last_n - the last N frames. Endpoints are a first/last-frame generation's bread and butter.
• specific - explicit comma-separated indices like 0,5,10,75. The mode for keyframes you've already identified.
• every_nth - a stride. Combined with num_frames you get "every 3rd frame, take the first 8 of those," which is the even-preview pattern.

Inputs are images plus mode, with num_frames (max 16), frame_indices, and step as the mode-specific params. The outputs are the quietly clever part: instead of a fixed set, the node exposes frame_1 through frame_16 and only activates the ones you actually select, so a two-frame grab shows two sockets and a 16-frame grab shows sixteen. It's the dynamic-input pattern that makes ComfyUI graphs stay readable instead of sprouting a wall of empty sockets.

Where it earns its keep

The three workflows this node exists for:

1. First/last-frame video - grab frame 0 and the final frame, wire them into a Wan or LTX first+last frame sampler. That's the single most common AI-video pattern, and it's two seconds of node work here.
2. Keyframe extraction - run a detection pass (or just know your shots), type the indices, and feed exactly those frames to a reference-aware model or a keyframe builder.
3. Preview contact sheets - every Nth frame through a grid node gives you a skim-able overview of a long render without decoding 400 frames.

There's a pleasant honesty in the mode naming here - the node isn't trying to be a shot-detection tool (the same pack's Cut Detective does that properly). It's a selector: you tell it what you want and it fetches it.

Gotchas

• specific mode wants indices within range. The node doesn't silently clamp out-of-range indices in all cases, so double-check your list against the batch length if you're scripting it from a report.
• every_nth with num_frames - remember num_frames caps the output count, step sets the stride. If you want "all frames every 3rd," set num_frames high; the cap of 16 outputs is the practical limit.
• The dynamic outputs can confuse saved workflows if the mode changes between saves. If a graph loads with sockets you didn't expect, it's usually the mode.

It ships in TrentNodes - ComfyUI Manager (search "Trent Nodes"), or clone + pip install -r requirements.txt. No models, no downloads, just indexing. It's the kind of node you won't think about until you've re-built the same slice-with-a-custom-node three times, at which point you'll be glad it's there.