MiniMax H3 Pixel Upscale Conditioning (Experimental)

Upscaling isn't just "make the pixels bigger" - in this pack's pixel route it's a genuine second diffusion pass, and a second pass has a sneaky failure mode: the reference images and keyframes that steered the first pass were encoded for the old, small canvas. Feed those to the refiner at the new resolution and you get conditioning that's spatially wrong, soft references, or detail that fights the upscale. The MiniMax H3 Pixel Upscale Conditioning node (experimental) exists to fix exactly that, and it fixes it the pragmatic way: it looks at the real dimensions of the upscaled images you give it, instead of guessing from a latent.

You reach for it between an IMAGE upscaler and the refiner. The upscaler produces bigger frames, this node rebuilds the scene's cached references at the target canvas and resizes the remaining eligible conditioning to match, and its outputs feed the guider and the image refiner. No fake latent, no whole-video VAE encode just to measure size - that's the design win here.

How it works

The scene's original pass cached its native Ref2VA reference data (assuming the source pass had cache_for_upscale on). This node restores that cache, re-encodes picture references at the target resolution using your H3 CLIP and video VAE, and synchronizes any remaining match-sized video/keyframe latents to the measured scale_x/scale_y between the upscaled images and the original canvas. Max-sized pictures deliberately keep their capped geometry, and the default motion-reference policy drops reference video while keeping paired audio - motion refs sized for the old canvas rarely survive an upscale usefully.

The one hard rule: your input images must already be H3-aligned - width and height in multiples of 32. There is no silent resize here. If your upscaler produced a fractional size (many do at "2.3x"), align it yourself before this node or it throws.

Inputs that matter

• images - the actual upscaled RAW frames, before USDU/refinement. Their width/height drive everything. Connect the same images to the image refiner.
• clip and video_vae - the H3 text encoder and video VAE used to rebuild the cached Ref2VA presentation at pass 2.
• state - current Upscale state from the adapter.
• method - interpolation for eligible cached video/keyframe latents; bilinear is the default and fine.
• motion_ref_mode - exclude_video_keep_audio (default) drops native video refs while keeping paired audio; the other two keep or resize them.

The optional prompt_override is worth knowing: leave blank to reuse the exact compiled generation prompt, or write a concise appearance/detail prompt so pass 2 doesn't re-trigger the original motion and camera instructions. tagged_references lets you override the cached reference line entirely; missing_cache decides whether runs with no matching cache fall back to text-only or hard-error.

Outputs

positive conditioning goes to a fresh Basic Guider for the second pass. images passes your upscaled RAW frames through unchanged to the refiner (in the example flow, USDU). You also get width/height, the actual compiled_prompt, a reference_cache_used boolean, and a status summary of target size, scale, and policy.

How to install

ComfyUI Manager (search "MiniMax H3 Context Loop"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/seitanism/ComfyUI-H3-Motion-Context-MultiRef.git
git clone https://github.com/ethanfel/ComfyUI-MiniMaxH3-Context-Loop.git

Restart ComfyUI. Beyond the two repos and your own H3 model/VAE files, the full example needs the extra external packs (DLSS5 and an H3-aware USDU guider fork) - this conditioning node itself just needs the source run and a CLIP/VAE.

Common issues

The error you'll actually hit is the alignment one: "Pixel target WxH is not H3-aligned." Resize to multiples of 32 before feeding this node, and don't wedge a resize between it and the refiner or the dimensions drift again. It also requires the complete RAW scene - trimmed or interpolated frames get rejected loudly, which beats silently wrong conditioning.