arkennemasis fal · MiniMax H3 Max Camera Controls · $0.05-0.16/s

One picture in, a camera move out

Most image-to-video nodes will happily pan and zoom if you ask nicely in the prompt, and then do something else anyway. This one takes a list of camera poses and follows it. Give it a still and three keyframes and you get a controlled orbit, arc or pull-back - the camera moving, the scene staying put.

That distinction is the point. The node's default prompt says it out loud, and it's worth reading twice:

"The same elements in the reference are rigid. Preserve every element exactly. The entire scene is frozen. Only the camera moves. no scene motion only camera motion"

You're getting a camera move, not an animation. Great for product shots, architectural reveals, turntable-style loops for a still, or establishing a 3D-feeling shot from a single image. Wrong tool if you want the subject to do anything.

Pricing is the H3 Max video ladder - $0.05/s at 480p, $0.08/s at 768p, $0.16/s at 1080p - with the note that fal was running 50% promotional rates on some H3 Max endpoints through the end of September 2026, so check what your node's badge says today rather than trusting a blog post.

The input that does the work

camera_trajectory - JSON, ordered keyframes, and the only genuinely fiddly thing here. Each entry has four fields:

[{"time": 0, "elevation": 0, "azimuth": 0, "distance": 1},
 {"time": 0.3, "elevation": 15, "azimuth": 45, "distance": 0.2},
 {"time": 0.8, "elevation": -90, "azimuth": 90, "distance": 1}]

time is the normalised position in the clip (0 to 1), azimuth is the horizontal angle, elevation is up/down, and distance is how far the camera sits from the subject. The first pose holds before its time, the last holds for the rest of the video, and signed full turns are respected with a cap of 32 turns of total azimuth travel - so you can loop the camera right around if you want.

The shipped default is a usable demo, not a recommendation: it swoops in close at 45°, then drops underneath and swings round. Replace it with two or three poses you actually want, and keep the intervals even unless you're deliberately accelerating.

The rest:

• image - the required first frame. Canvas follows it.
• prompt - present with that frozen-scene default. Blank leaves it out; that's usually correct.
• duration - seconds, default 5. Your trajectory's time values are fractions of this.
• resolution - 480P (default), 768P, 1080P. The 1080P tier is latent refinement from a native 768P source.
• seed (-1 = model picks), enable_safety_checker (on), and prompt_expansion_mode (disabled, fast, balanced default, quality - the last one can spend ~30s enriching your prompt, which is exactly what you don't want for a "nothing moves" instruction).

Guards: max_cost_usd ($20 default, 0 off), reuse_identical_run, extra_json, max_concurrent (1 by default).

Outputs

video (to SaveVideo, PreviewVideo, or the pack's Video Assemble), video_path, timings (how long fal spent on each stage - the honest place to look when a run felt slow), expanded_prompt, and info with the request id, estimate and fal's raw answer.

Install

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

Restart, then arkennemasis/fal/Video/MiniMax. .env next to the launcher:

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

Where people get burned

JSON is JSON. A trailing comma, a missing bracket, or counts wrapped in quotes and the run fails before it costs you anything - the node validates extra_json and its own JSON boxes before upload. If you edit it in a text editor and paste it back, check the shape first.

time is not seconds. It's the fraction of the clip. A pose at time: 5 on a 5-second video is meaningless; a pose at time: 1.0 is the last frame, which is where the hold begins.

Distance is the motion. A trajectory that changes only azimuth gives you a flat orbit. The pull-in/pull-out that makes these clips look expensive usually comes from moving distance as well.

Over-ambitious poses read as a glitch. Full turns in 5 seconds, or an elevation that flips the camera upside down mid-shot, produce something that looks like a render error. Signed full turns are supported - that doesn't mean they're tasteful.

The prompt fights you if you let it. With prompt expansion on quality, the model may rewrite "only the camera moves" into something livelier and give you scene motion you spent money trying to avoid. disabled is the safe setting here.