H3 Continuum Still Image Guide V3.7

The two-field node that makes the V3.7 guide work

This is the plumbing node of the H3 Continuum pack's V3.7 experiment: it takes an image and a frame number, wraps them into a guide object, and hands it to H3 Continuum Sampler V3.7. Nothing generates here - no sampling, no VAE, no model. Its whole job is to say "at this exact frame, this exact image should be in the video," in a format the sampler understands. If you've only ever seen it sitting orphaned in a workflow, that's normal; it's the prep step for something that lives downstream.

You use it when you want one specific frame of a long-form MiniMax H3 video to land on a specific picture. Think of it as a hard keyframe: "the shot at frame 156 should be this." That's a different ask than reference images (which steer identity throughout) or a Video Guide (which steers motion and framing across every chunk). This is a single point-anchor on the timeline.

How it works

Two inputs, one output:

• image - any IMAGE tensor, typically from a Load Image node. The sampler resizes it to your output latent geometry when it applies the guide, so don't overthink source resolution; just feed something clean.
• absolute_frame - an integer, default 0. This is the zero-based frame on the final visible 24 fps Continuum timeline, not a frame within a chunk and not your source video's frame numbering. The sampler then resolves it to whichever physical sampling group owns it and injects it as a Core minimax_keyframes entry on only that group - chunks before and after stay on their normal path.

The math is straightforward once you remember everything is 24 fps. With 3 chunks × 5 seconds you have 120 frames per chunk. Want the anchor 1.5 seconds into chunk 2? That's 1 chunk × 120 + 36 = 156. Ten-second chunks? 240 per chunk. The [0-5s]-style headers in your timeline prompt are a good sanity check that your frame lands where you think it does.

Output is a single guide socket (type H3_CONTINUUM_STILL_IMAGE_GUIDE). Wire it into the V3.7 sampler's guide input. It won't connect to the V3.6 sampler - that node has no guide socket, which is exactly why this node only makes sense paired with V3.7.

Read the fine print before you rely on it

The README is unusually honest here: Still Image Guide is Experimental and on Production HOLD. The underlying Core Add Guide path uses hard-anchor semantics - the image is treated as a hard target at that frame, so the video can snap abruptly at the anchor and the guide can redirect later motion. It is not a smooth transition control, and the pack explicitly says not to treat it as one. It's the kind of feature you test on a short 2×5s run with a fixed seed, verify the anchor frame looks right, and only then trust in a longer render.

Install

It ships in the H3 Continuum pack, so one install covers everything:

cd ComfyUI/custom_nodes
git clone https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum.git

or search "H3 Continuum" in ComfyUI Manager, then restart. No extra pip dependencies - the requirements.txt is empty, and the pack needs only what ComfyUI already provides plus the MiniMax H3 weights themselves. One gotcha that has nothing to do with this node but stops everyone: the H3 Community License excludes the US, EU, UK and South Korea from running the local weights, so make sure you're licensed to run H3 locally before you build a whole workflow around it.

Where people get confused

The absolute_frame value is the most common mis-set field, because it's easy to think in "seconds into the video" instead of frames, or in chunk-local frames instead of global ones. Multiply seconds by 24, add the chunk offset, zero-based. And don't expect subtlety: the guide is a hard anchor. If your result looks like the video jerks to the image and then carries that influence forward aggressively, that's the documented behavior, not a bug in your settings.