TensorScale MiniMax H3 Video Reference (Soze)

The sibling of the MiniMax H3 video node, and the one to pick when you have source material instead of just a prompt. Where the plain H3 node conditions on first/last frames, this one is true reference-to-video-and-audio: nine image slots, three video slots, three audio slots, and a text field for raw URLs. It's the most reference-rich node in the TensorScale family, and the one with the most ways to blow past the API's size limit - so read the gotcha section before you wire a big video into it.

The inputs

Core set is the same as the sibling: prompt, duration_seconds, aspect_ratio (with the same aspect_ratio_override escape hatch), seed. Then the references:

• reference_image_1 through reference_image_9 - images that define characters, settings, objects.
• reference_video_1 through reference_video_3 - motion/style references from existing clips.
• reference_audio_1 through reference_audio_3 - the voice-cloning / native-sound inputs. A short clip of a voice here is how H3 makes the speaker sound like that person.
• reference_uris - a multiline text field where each line is type|https://... (or a bare URL whose extension implies the type). Anything pasted here is appended after the socket references, so it's the escape hatch for material you have as a public URL rather than a local asset.

The prompt is where you wire it together and describe the sound: "The narrator in reference_audio_1 introduces the hero from reference_image_1, who runs through the market in reference_video_1." Yes, the prompt is the mix console - H3's audio comes from the same generation pass, not a separate TTS node bolted on later.

How it works

TensorScale runs /v2/MiniMax-H3/ref2va synchronously with a model-scoped key (TENSORSCALE_API_KEY_MINIMAX_H3, falling back to TENSORSCALE_API_KEY). Media is inlined as base64 data URIs by default, and this is where the 10 MiB body cap bites hardest in the whole pack - a video reference inlined as base64 will exceed it almost immediately. The tooltips on the video slots say it plainly: turn on use_fal_upload, or feed raw URLs through reference_uris. use_fal_upload pushes the connected media to FAL's CDN, which means it needs a working FAL_KEY too.

Outputs: video, video_path, request_id, config, status - the standard TensorScale set.

Install and key

From the Soze pack ("Quality of Life Nodes for ComfyUI"), via ComfyUI Manager or:

cd ComfyUI/custom_nodes
git clone https://github.com/SozeInc/ComfyUI_Soze.git
pip install -r ComfyUI_Soze/requirements.txt

Restart, then set TENSORSCALE_API_KEY or TENSORSCALE_API_KEY_MINIMAX_H3 (environment variable or .env in the repo root). Don't commit keys.

Common issues

• Video references silently fail. They blew past the 10 MiB cap when inlined. Use use_fal_upload or reference_uris with a public URL - not optional for video.
• Audio reference doesn't clone the voice. The audio needs to be clean and short; a noisy clip gives a noisy voice.
• Key scope errors. This endpoint needs the MiniMax scope; a generic key or the wrong model-scoped variable won't authenticate.