H3 Keyframes (any position)

Stock H3 conditioning lets you anchor a clip at the first frame, the last frame, or both. This node throws the middle open. H3Keyframes builds MiniMax-H3 conditioning with keyframes at arbitrary positions in the timeline - connect an image to image_1, give it a position, and that picture becomes a fixed point the video passes through halfway through, or at frame 300, or wherever you say. This is how you get a clip that starts on one face, turns to the camera at the midpoint, and ends on a wide shot - dictated by three stills instead of hoping the text does it.

It's the conditioning heart of the pack's separate H3_Keyframes.json workflow, which is a different job from multishot chaining: this is one clip, anchored at chosen positions, with per-anchor condition strength available via H3ConditionStrength.

The inputs that matter

• clip, vae - the encoder and VAE for text and image encoding.

• prompt - the clip's text.

• width, height, length - render size and frame count (length on H3's 17k+5 grid; 243 is ~10.1s, 362 is the trained max ~15.1s).

• positions - the string that makes this node special. One entry per connected image, in order. A % value is a fraction of the clip: 0% = first frame, 50% = halfway, 100% = last. A bare number is an absolute frame index. You can also give ranges - 2-5 or 15%-30% - and ranges can run descending (30%-20%), which reverses that section of the image batch. Example from the tooltip: 0%, 50%, 100% anchors start, middle and end.

• image_1 … image_6 - the anchor stills. When six isn't enough, images_batch takes a whole batch of anchors (every frame is one anchor, in order, appended after the individually-connected slots) - say, several frames at each end to pin complex motion, or frames lifted from a source video.

Outputs: positive (the CONDITIONING, into your sampler), latent (the encoded latent), and info (a string describing what got anchored - read it; it'll tell you if your positions parsed the way you meant).

How it works

Each anchor is encoded to a latent and placed at its temporal coordinate in the conditioning, the same mechanism H3 uses natively for first/last frames - just extended to any position. The position parsing (0%, 50%, ranges, descending ranges) was contributed upstream by @viralesveras, and it's the kind of syntax you'll either love instantly or need one look at the info output to trust.

The honest warning is in the tooltip: large numbers of keyframes massively increase VRAM. Every anchor is a latent row that rides through every sampling step. Six anchors is fine; forty frames from a source video will want a big card. Plan the keyframes you actually need, not the ones you can supply.

Install

Part of the H3 Multishot pack:

cd ComfyUI/custom_nodes
git clone https://github.com/jlucasmcrell/ComfyUI-H3-Multishot

or ComfyUI-Manager → H3 Multishot. Restart; ComfyUI v0.30.0+. No extra dependencies.

The usual model install applies - checkpoint in models/diffusion_models/, encoder + VAEs in text_encoders/ and vae/. And the pairing rule from the rest of the pack applies here too: this is ref2va territory (reference rows), so if the anchors aren't holding, check your encoder's vision sidecar is actually loaded before you blame the keyframes.