H3RelayInternalLoadFrames

Not every node in a pack needs to be clever. Load Frames takes one input - a video_path string - decodes the video into an IMAGE stack, and returns it. That's the entire interface. It's the kind of node you scroll past a hundred times before you stop and wonder what it's actually for.

It's for feeding an existing video back into the pipeline as pixel frames. In H3 Relay's world that's mostly one job: pulling in a prior delivered segment so a continuation or enhancement step can condition on real frames rather than abstract state. The context-loop machinery needs actual pixels from the previous shot when it builds continuation conditioning, and this is the node that turns a path on disk into those pixels.

Why it's internal

Notice the video_path input is a string, not a file browser. That's a deliberate tell that this node isn't meant for humans - it's wired by the graph-expansion machinery with a path computed from the run's disk records. When a public node expands and needs "the frames of the previously accepted segment," it builds a Load Frames node and feeds it the path directly from state. You'll see it appear and disappear as graphs expand, and that's normal.

If you are reaching for it by hand, there's a plain-stock alternative that's usually the better call: ComfyUI's own video-loading nodes with a file picker. Load Frames is a convenience for paths you already have programmatically - for example, if you're writing a custom workflow and you know exactly which relative path inside ComfyUI output you want, this node accepts it directly. The input is the standard STRING socket, so you can also wire a path from another node rather than typing it.

The output is a plain IMAGE tensor - a batch of frames you can feed into anything that eats images: a VAE encode, an image-conditioning node, a reference-image socket, you name it.

The honest take

If you never manually place this node, nothing breaks. It's plumbing. Its existence is worth knowing about for one reason: when you're debugging why a continuation shot looks wrong, the question "wait, which frames did it actually feed back in?" points at this node, and now you know what the video_path input is doing and how to check it. The path resolves relative to ComfyUI's output directory, so if you're inspecting a graph, that's where the file lives.

Install the pack the usual way - ComfyUI Manager search H3 Relay, or git clone https://github.com/akatz-ai/h3-relay.git into custom_nodes, restart. Needs ComfyUI 0.32.0+ and FFmpeg on PATH; the model files in MODELS.md are what actually make the pack do anything, and the MiniMax H3 community license geofence (US/EU/UK/Korea) applies to the H3 side of things. No extra Python deps for this or any node in the pack.