Sample Video Frames (Images)

You've got a video, and you want a handful of its frames as images - but not just any frames, and not just as an image batch: you want them aligned to timestamps, as a list, with the <Picture N> references formatted for pasting straight into a video-model prompt. That's the oddly specific job this node does, and it does it better than the generic "extract frames" loops most workflows used to hack together.

What it does

Feed it a Core VIDEO input and it samples presentation-timestamp-aligned frames, then hands you the same frames in several forms at once: an image batch (IMAGE, chronological, good for direct sampling), an images list (IMAGE, one entry per frame, independently mappable), and a stack of text outputs - timestamps, timestamps text, timeline text, and structured timeline text - all formatted consistently and ready to drop into a prompt.

The default timeline_style is "H3 alignment prefix," which is a tell: this node was built for the pack's MiniMax H3 / Qwen video-reference workflows, where the prompt references images as <Picture N> entries with timestamps. The index_offset input exists so you can reserve the first indices for earlier reference images and let this node's frames start later in the numbering. If you're not on an H3 workflow, the "timestamps only" or "zero-based indexed" styles give you the same data in plainer form.

The settings that matter

• sampling_strategy - codec keyframes (default) keeps only actual codec keyframes (low redundancy); uniform PTS spaces frames evenly across the video; focused PTS groups frames around specific moments.
• focus_areas (0–3) + focus_one/two/three (0–1 each) - focused mode splits the video into up to three parts and each focus value says where in that part the frames cluster (0 = early, 0.5 = balanced, 1 = late). Want the interesting motion at 70% through the clip? Set focus to 0.7.
• maximum_frames (default 16; 0 = keep every eligible frame) and minimum_spacing_seconds (default 0.25) - together they cap output and enforce a floor between frames, so you don't get a burst of near-duplicate frames.
• include_zero_time (default on) normalizes the first visible frame to position 0 - the timestamp reference everyone expects.
• keyframe_stride - keyframe mode only; keep every Nth keyframe before spacing and count limiting.
• timestamp_format and timeline_style - cosmetic, but they set exactly how the text outputs read.

Outputs you'll actually wire: image batch into a sampler, the timeline text / structured timeline text into a prompt, and video runtime when you need the active duration (including any trim) in seconds.

Install

Part of ComfyUI-UtilsCollection:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-UtilsCollection

Restart, or install via ComfyUI Manager (search "UtilsCollection"). No model downloads - this is pure frame extraction. Requires Core's VIDEO type, which means a reasonably recent ComfyUI.

Gotchas

Keyframe mode depends on the codec actually having keyframes; highly compressed web sources may yield fewer than you expect - bump maximum_frames or switch to uniform PTS if your picks look sparse. And remember the pack-level housekeeping: UtilsCollection re-registers node IDs from several older packs, so accept ComfyUI's "replace node?" prompt when loading older workflows and you'll be on the current implementation.