TensorScale MiniMax H3 Video (Soze)

MiniMax H3 is one of the few frontier video models that comes back with a real audio track baked in - speech, foley, and music generated alongside the picture. This node serves it through TensorScale, an inference service that works differently from the FAL nodes: instead of polling for a job, it holds the connection open and streams the finished clip straight back. One POST, one file, no job ID juggling.

The inputs that matter

The required set is small: prompt, duration_seconds, aspect_ratio, and seed. The prompt describes the whole scene including what it sounds like - that's the H3 twist. "A train pulls into a rainy station; announce the arrival over the PA" is the shape of a prompt that uses the model's full capability.

• duration_seconds (1–60) gets snapped server-side to H3's native 24 fps / 17n+5 frame boundary, so don't expect a clip of exactly 8.0 seconds.
• aspect_ratio runs from 4:1 ultrawide to 1:4 tall; 16:9 renders at 1344x768, 1:1 at 768x768. A free-form aspect_ratio_override (e.g. "2:1") replaces the dropdown if you need something in between.

The optional framing inputs are where H3 gets interesting. Leave both empty and you get pure text-to-video+audio. Connect first_image and it becomes the opening frame (stretched to the canvas). Add last_image too and you've bookended the clip - H3 animates between your two endpoints, the first/last-frame (FL2VA) mode. URLs in first_image_url/last_image_url override the sockets.

Then the tuning knobs, all optional: num_inference_steps (native profile is 50), flow_shift (video scheduler, profile wants 12.0), audio_flow_shift (audio scheduler, profile wants 3.0). The defaults match the service profile - leave them alone until you have a reason not to.

How it works

TensorScale's model-scoped API means your key must carry the minimax-h3 scope, and it's read from the env var named by api_key_env (default TENSORSCALE_API_KEY_MINIMAX_H3, falling back to TENSORSCALE_API_KEY). The node POSTs to /v2/MiniMax-H3/fl2va and the socket stays open for the whole generation - the timeout widget (default 1800s) is a read timeout, so long clips legitimately take minutes.

The critical gotcha: the request body is capped at 10 MiB, and connected images are inlined as base64 data URIs by default. A couple of hi-res PNGs can bust the cap. If you're feeding frames, flip use_fal_upload on - it pushes the media to FAL's CDN and sends URLs instead (which means that toggle needs a working FAL_KEY even though the generation is TensorScale).

Outputs: video (the VIDEO object - preview or save it), video_path (where the mp4 landed), request_id, config (a summary string of what ran), and status.

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 the model-scoped TENSORSCALE_API_KEY_MINIMAX_H3) as an environment variable or in a .env file in the repo root. Never commit it.

Common issues

• "Request body is X MiB, over TensorScale's 10 MiB limit." Your reference frames are being inlined. Turn on use_fal_upload or use URL widgets.
• Clip length isn't what you asked for. H3 snaps to its frame grid - 8.0s comes back as whatever 17n+5 lands nearest. That's the model, not a bug.
• use_fal_upload fails. It needs FAL_KEY configured even though this is a TensorScale call - the upload hop goes through FAL's CDN.