H3 Continuum Sampler V3.7

If you already run V3.6, you already know this node

H3 Continuum Sampler V3.7 is the current release path of the pack, and the honest summary is: it is the V3.6 sampler with exactly one extra optional socket. Same chunked continuation, same Masked AV behavior, same Run Storage, same production defaults - the pack is explicit that V3.7 does not change any V3.6 defaults. The one new thing is a guide input that accepts a Still Image Guide, and that one new thing is worth reading about before you plug it in, because it's genuinely experimental.

So when do you reach for this node instead of the V3.6? If you never touch the guide socket, it's a straight upgrade in name only - same outputs, same wiring, same workflow. If you want to pin one specific frame of your long-form video to a particular image, this is the node that lets you do that while keeping everything else identical.

How the guide works

The guide input takes a special object type produced by the pack's H3 Continuum Still Image Guide V3.7 node - you don't build the guide inline, you make it with that two-input prep node and wire it across. That guide packs an image plus an absolute frame number on the final 24 fps timeline. When the V3.7 sampler runs, it resolves that absolute frame to whichever physical sampling group owns it and injects the image as a Core minimax_keyframes entry on only that group. Chunks before and after it sample normally.

Here's the part the README is blunt about: Still Image Guide is on Production HOLD. Core's Add Guide uses hard-anchor semantics - the image is treated as a hard target for that frame, so it can cause an abrupt trajectory change at the anchor and can redirect later motion. It is not a smooth "drift toward this image" control. If your mental model was "lens the scene toward this reference over a few seconds," disabuse yourself now. If your actual goal is "frame 156 must look like this exact shot," that's what it's for.

Inputs and outputs

Everything from V3.6 is here, unchanged: model, clip, video_vae (image-conditioning encode only - Continuum never decodes with it), sampler, sigmas, sequence_prompt, chunks, chunk_seconds (5 default, 5–15 validated), width/height (multiples of 32), continuity, base_seed, audio_continuity, continuation_backend, run_storage/run_name/reroll_from_chunk, and the same optional first_frame, last_frame, reference_image_1–3, reference_video_1, driving_audio, and reference-audio inputs. Plus the new one:

• guide (optional) - the Still Image Guide object from the guide-prep node. Leave disconnected and you're running the V3.6 production path exactly.

Outputs are identical to V3.6: video_latents and audio_latents (feed to Core VAE Decode), assembly_plan (into H3 Continuum Assemble + Seam), status (a string - read it, it reports resolved transport and chunk reuse), driving_audio, and refine_context (for Second Pass / Hi-Res Fix). A word of advice: if you're comparing guide-on vs guide-off behavior, fix your seed, because an abrupt anchor plus a new seed is two variables at once.

Install and setup

Same as every node in this pack - ComfyUI 0.32.0+:

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

or search "H3 Continuum" in ComfyUI Manager. Restart after. No extra pip dependencies (the requirements.txt is empty), but you need the MiniMax H3 weights - and remember the H3 Community License excludes the US, EU, UK and South Korea from running them locally, so verify you're in a permitted territory before you start.

The trap to avoid

Treating the Still Image Guide as a continuity tool. The whole point of V3.7's "Stage 2" positioning is that the guide exists to be tested, not to be your default workflow. The pack's own correctness gates validated that the guide maps to the right group, keeps ownership, and isolates non-owner chunks - and still kept it Experimental because the anchor semantics are harsh. Use it when you need a hard keyframe; if you want smooth transitions, that's what driving the sampler with references and strong continuity prompts is for. Don't expect the guide to fix drift it isn't designed to fix.