MiniMax H3 Reference (URL) to Video

If Reference to Video is the full-strength version of this pack, this is the streamlined one: same omni-modal conditioning - images, videos, audio, all feeding one generated clip with native sound - but the references come in as multiline URL text instead of tensors, so nothing ever gets uploaded by your machine. You've got image and video URLs already hosted, or an mm_file:// reference from an earlier run, this is the node you reach for. Your reference assets are local files on disk, you want the tensor sibling instead, because there's no file browser here.

How it works

Three text fields, one URL per line. The node strips and splits each field, filters out blank lines, and drops the URLs straight into the generation request's content array - images first (role reference_image), then videos, then audio. No /v1/files/upload call at all, which removes both the upload step and its failure modes. Then it's the pack's standard routine: POST to /v2/video_generation, poll every 10 seconds until the task completes, download the result as a VIDEO output.

The reference numbering scheme is the same as the tensor node and it's the thing you must get right: in the prompt, refer to inputs by per-type name and order - Image 1..Image 9, Video 1..Video 3, Audio 1..Audio 3. URL order in the text field is reference order, so line 1 of the image field is Image 1, and so on. Supported URL forms: public URLs, mm_file://{file_id}, or data URIs.

One limit that carries over from the tensor version: at least one image or video URL is required, and audio URLs can't be used alone. The video budget (each 2–15s, 15 seconds total) applies here too, though since the node never inspects the files, it leans on the API to enforce duration - one more reason to pre-check your URLs.

The inputs that matter

• prompt - required, and where you reference the inputs by name (Video 1, Audio 1, ...).
• reference_image_urls - multiline, up to 9.
• reference_video_urls - multiline, up to 3 (advanced, collapsed by default).
• reference_audio_urls - multiline, up to 3 (advanced).
• ratio - defaults to adaptive, full list available; unlike text-to-video there's no ban on adaptive here.
• resolution, duration, watermark - 768P/2K, 4–15s, and the AIGC watermark toggle. The seed input does nothing, as with every node in this pack.

Install

Same pack, same two routes. Manager → Install via Git URL, paste https://github.com/meta-sota/ComfyUI-MiniMaxH3-API.git, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/meta-sota/ComfyUI-MiniMaxH3-API.git
cd ComfyUI-MiniMaxH3-API
pip install -r requirements.txt

Just aiohttp from pip; needs a recent ComfyUI (author tests 0.30.0) because the pack is built on the native V3 node API. No model downloads, no GPU - the Key node provides credentials.

Common issues

• "At least one reference image or video URL is required." - audio-only isn't allowed.
• Dead or unreachable URLs - there's no local fetch, so the API is where a broken link surfaces, usually as a clear HTTP error. Test the URL in a browser first.
• URL order mix-ups - the model can't reorder your references; if the output ignores Image 2, check that line 2 of your text field is actually the image you think it is.
• Duration surprises - since the node can't measure the videos, an over-budget clip fails at the API instead of at validation. Keep reference videos under the shared 15-second total.