arkennemasis fal · MiniMax H3 Max Insert Video · $0.05-0.06/s

Fix the middle of a clip, not the whole thing

Insert Video takes an existing clip and injects a newly generated scene into the middle of it, then returns to the original footage. You guide the new scene with a prompt, reference images, reference videos, or any combination.

That's a specific and very practical job. The take is good except for eight seconds where the handoff was wrong. The client wants a different product in frame for one beat. You want a short cutaway that isn't in the footage. Regenerating the clip is the expensive, lossy answer; inserting a scene is the surgical one.

The billing matches the ambition: $0.05/s of generated scene at 480p, $0.06/s at 768p - and only the new scene duration is billed, not the full output. The first 4,096 reference tokens are included, with extra reference tokens at $0.02 per 1,000, prorated. So a 5-second insert into a 30-second take is a 5-second bill, roughly a quarter.

Fair warning before you start: this is a new endpoint with essentially no community track record - you won't find threads about it, because there aren't any yet. Read the failure modes below as mechanism, not folklore.

How it works

Your source video and any references upload to fal's CDN (free - uploads never bill), one request goes to minimax/h3-max/insert-video, and the finished clip lands in output/fal/minimax_h3-max_insert-video/. The node returns a lot of scalars because this endpoint tells you exactly what it did - frame count, dimensions, the injected duration and where it landed.

The inputs that matter

• video - the source clip you're editing.
• start_time and resume_time - the splice window, in seconds. Defaults are 1.625 and 1.635, i.e. essentially the same instant just past the minimum: start_time has a floor of 1.625 and resume_time a floor of 1.635. start_time is where the new scene goes in; resume_time is when the original footage resumes. Pushing them apart means you're replacing that stretch of source; leaving them together means you're inserting without removing anything. This pair is the whole edit.
• duration - the length of the generated scene, a float between 5 and 13 seconds, default 5. This is what gets billed, so it's the number to watch.
• prompt - what the new scene is. Optional; empty leaves it out entirely.
• reference_image_1…6 and reference_video_1…3 - the guides for the generated scene, each socket optionally carrying a batch, sent in socket order (first picture on socket 1 is number 1). A couple of reference frames from your own take is the cheapest way to keep the insert looking like it belongs.
• resolution - 480p or 768p, default 768p. Nothing higher on this endpoint.
• color_match - on by default, and I'd leave it on. It's what stops the generated scene from reading as a different camera, different grade, different shoot.
• enable_prompt_expansion - on by default. It costs time and can enrich a thin prompt; turn it off if you want your wording used verbatim.
• seed - -1 leaves it to the model.

And the standard guards: max_cost_usd (default $20, checked before upload so an accidental 13-second scene at the wrong settings gets stopped, not billed), reuse_identical_run, extra_json, max_concurrent (1 by default across the whole graph).

Outputs

video (the whole clip with the insert in place - straight to SaveVideo, PreviewVideo, or a Video Assemble), video_path, and a useful pile of metadata: frame_count, width, height, start_time, resume_time, duration, injected_duration (what the scene actually came out as, which may not be exactly what you asked), seed (the one actually used), expanded_prompt, timings, source, and info with the request id and cost estimate.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis
pip install -r comfyui-arkennemasis/requirements.txt

Restart ComfyUI, then look under arkennemasis/fal/Video/MiniMax. One line in the .env next to run_nvidia_gpu.bat:

FAL_KEY=your-key-from-fal.ai/dashboard/keys

There's no key field on the node by design - that's what keeps a key out of your saved workflows.

Where people get burned

Replacing when you meant inserting. With resume_time set barely past start_time, the source keeps playing right after your generated scene. Spread them apart and that stretch of source is gone. Decide which one you're doing before you spend the money.

Expecting a frame-perfect handoff. This is a generative splice: the model generates a scene that should sit in the neighbourhood, not a compositing operation with alpha edges. Trust color_match tech, and check the returned start_time, resume_time and injected_duration rather than assuming your requested numbers came back exactly.

A 13-second insert is not a cheap experiment. Five seconds is the minimum billed scene, and it scales up: 13 × $0.06 ≈ $0.78 at 768p, per attempt. Draft your intent at 480p on a short scene first.

Reference order is the only addressing scheme. The prompt refers to references by number, and numbers follow socket order with batches flattened - first picture on socket 1 is 1. If your prompt says "use Image 2" and you get something baffling, that's where to look.

No track record means no shortcuts. Nobody has posted the gotchas yet. Do one small paid run before you build a workflow that depends on this node, and read the info output when it finishes.