🔍 H3 Relay · Latent Upscale

The other seven nodes are all about time - where one segment ends and the next starts. This one is about resolution, and it's the only node in the pack whose name doesn't start with 🔗. It takes the packed H3 AV latent, unpacks it, runs the temporal blocks through a learned 3D upscaler, and repacks. Zero denoising, no frames added or removed, audio carried through untouched. It is a pure geometry operation on the latent.

Two reasons to reach for it. First, the obvious one: you're rendering at H3's native resolution and want more pixels without a full second sampling pass. Second - and this is the interesting case - it's the intermediate step in the standard production route: sample at native → upscale the latent → sample again at high resolution, which is the video equivalent of a hi-res fix. The upscaling essay covers why that second low-denoise pass is what adds coherent detail rather than just pixels; this node is the step that gets you there cheaply.

The pack takes no credit for the upscaler itself. The learned model comes from a separate community package, Comfyui_Minimax_h3_latent_Upscaler by LBH-123-AI, which you install alongside. This node won't work without it - and the error it throws names the repo and the weights URL, which is better than most cross-pack dependencies manage.

The input that will bite you

Most of the schema is polite: mode picks how you specify the target (a scale multiplier, an explicit width/height, or a megapixels budget), and only the matching parameter is read. align must stay 32 - that's a hard requirement of the H3 latent upscaler, not a preference. precision (bf16 default) and device are exactly what they look like; force_unload pushes the upscaler back to CPU when it's done, which is worth turning on if a sampler follows, and worth leaving off when you're chunking because it'll thrash.

Then there's chunks, and this is the one to understand before you touch it. chunks: 1 runs the whole segment in one pass - fastest, hungriest, and the only setting that matches the upstream model's inference exactly. Raising it splits the segment along the time axis to fit in VRAM, and it changes the picture: the upscaler's attention is volumetric in 3D, so splitting it severs the cross-chunk temporal context. overlap (default 5, matching the model's temporal kernel) blends the borders and softens the break - it does not cancel it. Measured per-frame MAE is around 5.62/255, which is visible. The author's own framing is that the unit tests lock down the splicing maths, not "different chunk count, same picture". So: leave chunks at 1, raise it only when you're out of VRAM, and re-inspect your seams after you do. The maximum legal chunk count is bounded by the overlap constraint and printed in the report - exceeding it errors rather than quietly adjusting.

Output: latent, plus report with the real numbers.

The relay-specific rule

If you're using this mid-film, keep the continuation contract in the native domain. That means: Latent Save taps the first pass, before the upscale; the Copy Bridge's conditioning output is connected only on the native-resolution sampling pass, because minimax_keyframes rows are computed against the target grid and a resolution change makes the anchor row count not match - you get a shape mismatch from inside the model. The latent-side pin (prefix copy + noise mask) is grid-agnostic and still works on the second pass, so you're only giving up the framing half of the pin. Also: upscale once per chain. If your sampler already contains a built-in upscale step, this node is doing the same job a second time, on the same latent.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ZenHG/ComfyUI-H3-Latent-Relay.git

# only needed for the upscale node:
git clone https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler.git

Then download the weights from https://huggingface.co/LBH-123-AI/Minimax_h3_latent_upscaler into ComfyUI/models/latent_upscale_models/ and restart the backend. These are H3-specific 24-channel latent upscaler weights, not an ESRGAN-family pixel model. Dropping an ESRGAN checkpoint in there won't work and won't tell you why. ComfyUI Manager → search h3-latent-relay for the main pack; the other seven nodes run fine without any of this.

Where it goes wrong

Nothing shows in the model_name dropdown means the weights aren't in latent_upscale_models/, or you didn't restart after putting them there. A scale below 1 errors from upstream - this node only enlarges. And if your seam was already teetering, chunking is a good way to find out: re-check the joins after any chunks > 1 run, because you've changed the picture, not just the speed.