Images to Video Timeline

The modern image-to-video models - MiniMax H3 especially - don't just take a pile of reference images; they want to know when each image matters. H3's prompt format references images as <Picture N> entries with timestamps. If you've ever counted seconds by hand to place eight storyboard frames across a 12-second clip, UC_ImagesToVideoTimeline is the node that does that counting for you, in several output formats at once.

Feed it ordered images and a duration, and it emits per-image timestamps, a full timeline prompt ready for a model, a plain text line, and a combined batch - everything you need to wire a storyboard into an H3 or similar prompt without a calculator. It's the bridge between "I have frames" and "here's a video prompt with a real timeline."

How it works

The core behavior is even spacing: with focus_areas at 0, images are distributed evenly across duration. Set focus_areas to 1–3 and the node splits the video into parts, letting focus_one/focus_two/focus_three (0 = early, 0.5 = balanced, 1 = late) cluster images within each part - so you can pack more frames into the moment where the action is. last_image_is_final pins the last frame to the end of the video, and index_offset skips Picture numbers (useful when your prompt already uses some). timestamp_format (default 00.000s) controls how times look, and timeline_style picks the text shape - default H3 alignment prefix, or H3 pictures, indexed, or timestamps only.

resize_images (default on) normalizes all images to the first image's size for a clean batch; turn it off and the batch goes black, with each image kept at its own size on separate outputs.

The outputs that matter

• timeline text - the H3-style prompt text, ready to paste into a video prompt.
• structured timeline text - video length plus Picture timestamps in a structured form.
• timestamps - one timestamp per image, for wiring into per-image conditioning.
• images / image batch - the ordered frames, either individually or batched.
• video runtime - the duration in seconds, handy as a float downstream.

Inputs are duration, focus_areas, the focus sliders, plus the image_inputs autogrow sockets where you add frames in order (each image in a batch gets its own timestamp).

Installing it

ComfyUI Manager → search "ComfyUI-UtilsCollection", or:

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

Restart. Light deps (opencv-python, typing-extensions), no models, no API keys. Newer-API pack - keep ComfyUI updated.

The gotcha: those focus_* values are placement hints in a prompt, not hard frame anchors - the model treats Picture timestamps as instructions, so results stay stochastic and a "late" cluster won't be surgically precise. And if you see a black image batch, resize_images is off and you're expected to use the individual images outputs instead. That's by design, not a bug.