MiniMax H3 references (Reanimator)

MiniMax H3 references (Reanimator) is reference-to-video for H3: you supply up to nine images and optionally one video, and the model treats them as material - a character, a style, a movement - rather than as frames of the output. ComfyUI's own MiniMaxH3ReferenceToVideo takes each reference on its own socket (ref_image_1…ref_image_9, ref_video_1). This node takes the whole list as newline-separated file names and calls it once.

What makes it different from the keyframe node

Keyframes are the skeleton; references are the casting. With references the prompt names them positionally - image 1 is <Picture 1>, image 2 is <Picture 2> - so order is a parameter. Same for the video: <Video 1>. Nothing gets described automatically; the images are read by the node, your prompt does the talking.

The mechanism is small and worth knowing because it explains two behaviours. The names are split on newlines and loaded one at a time, each at its own resolution - a batch would force every reference onto a single size, which would be wrong for a mix of a face close-up and a wide reference. Each is EXIF-rotated (phone photos lying sideways get fixed) and converted to RGB. Then the optional video is pulled apart: its frames are resampled to H3's 24fps read rate by picking indices - a 30fps phone clip handed over untouched would play about 25% slow inside the model - and its audio rides along as the reference soundtrack when there is one.

Then it's one call into ComfyUI's reference node: positive and latent back out - the conditioning to your guider, the latent to SamplerCustomAdvanced's latent_image.

The inputs a beginner touches

• clip - the Qwen3-VL text encoder. H3's encoder is huge; the shipped variant is ~15.7GB of nvfp4, which on Ampere is emulated rather than accelerated, so prompt encoding is slow in a way that's easy to mistake for a broken render.
• vae and audio_vae - both required, and that's not an oversight. H3 generates picture and stereo sound jointly, so there's no silent path: no audio VAE, nothing runs.
• prompt - wired, not typed. This node declares the prompt as a forced input, so you connect a string node to it. Surprises people.
• width / height - 1344x768 by default, both multiples of 32.
• length - 124 frames by default. H3's latents want 17k+5 frames at 24fps, so useful values are 39, 56, 73, 90… 124 (7×17+5, about 5.2s). The pack's template floors this at 39 and notes that anything past your last reference is GPU time against footage you won't use.
• image_paths - one bare file name per line, exactly as it sits in ComfyUI/input. No subfolders, no drive letters. The node checks this and refuses anything that isn't a plain name in the input directory.
• ref_image_size - the only enum, match or max. Every H3 reference template ships match.
• video - optional. Leave it unplugged for pure image references, or text-only if you plug nothing in at all.

Installing it

Same pack as its siblings, which is the Reanimator Bridge - local generation for reanimator.app - though the nodes work standalone:

cd ComfyUI/custom_nodes
git clone https://github.com/unluckyandlucky/comfyui-reanimator
pip install cryptography Pillow

Manager was still "pending registry review" when the README was written, so clone it or drag the ZIP into custom_nodes and restart. Category is Reanimator. It imports ComfyUI's MiniMax H3 reference node at run time, so an older ComfyUI will show the node and fail when you queue.

Where people get burned

You need the ref2va checkpoint, not the keyframe one. References and keyframes are different model files: this pipeline wants the ~21GB ref2va int8_convrot checkpoint plus a 4-step turbo LoRA, where the keyframe graph uses fl2va with an 8-step LoRA. Downloading one does not get you the other.

Refs go in ComfyUI/input, nothing else. Paste refs/hero.png or an absolute path and you'll get "Not a file in ComfyUI's input folder" - that's the node refusing path traversal, not a bug. Ten references raises a hard error at nine.

Shape mismatch. References keep their own size; your canvas still has to be 32-aligned, and nothing here resizes the reference video's frames. If that footage is a different shape from your output, normalise it upstream - FitToGrid over its frames is the blunt way.

The 24GB tier is real. Pruned weights, Turbo LoRA, int8 - and the pack still calls 24GB the "full" tier on a 3090 that measured around 516 seconds for a 5-second cold clip. Convrot int8 needs ComfyUI v0.27.0 or newer for the native ops.

And the licence shapes who this is for. H3's community licence excludes the US, EU, UK and South Korea from its Applicable Territory, outputs included. Know it before the download.