arkennemasis fal · MiniMax H3 Max Styles Low Poly · $0.08/s

The look Gen Z calls retro and the rest of us call 1997

This endpoint generates 768p video with audio in a retro low-poly 3D style, from a text prompt or an optional first frame. Prompt in, faceted characters and hard-edged environments out, no prompt gymnastics required - the style is applied automatically.

It's a narrower use case than the other style nodes, and worth being honest about: low-poly is a specific aesthetic, and if you want it, this saves you a pile of failed attempts where a general model keeps smoothing everything into realism. If you're making something that needs to look like a PS1 game, a 90s flight sim, or a stylised product teaser, this is the fastest route there. It's also genuinely good for previsualising a 3D scene before you model it - the motion reads clearly when everything is blocks.

Price is fixed for the tier: $0.08 per second at 768p with audio included, five seconds minimum-ish at $0.40, fifteen at $1.20. Durations run 5–15 seconds.

Inputs that matter

• prompt - scene, action, camera, sound. Remember the style is automatic; describing "low poly, PS1, flat shaded" is redundant and can push the model around for no reason.
• image - an optional first frame, and the tooltip's steer is specific: use low-poly or retro 3D artwork for the most consistent look. Its aspect ratio also becomes the output canvas, so a square game-art frame gives you a square clip.
• duration - seconds, default 5, supported range 5–15.
• aspect_ratio - 21:9, 16:9 (default), 4:3, 1:1, 3:4, 9:16. You'll notice 4:3 is in the list: that's the one that sells the era, and it's ignored when you supply an image.
• seed - -1 by default; pin it for a repeatable take.

The shared fal-node guards are all here too: max_cost_usd (default $20, checked before upload), reuse_identical_run (no double charge for identical inputs, even after a restart), extra_json (raw JSON for anything the widgets don't expose), and max_concurrent (default 1 - one paid fal call at a time across the whole graph).

Outputs

video (into SaveVideo, PreviewVideo, or Video Assemble for a multi-shot cut), video_path, seed (what was actually used) and info with the request ID, cost estimate, saved files and fal's full response.

Install

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

Restart ComfyUI and it's under arkennemasis/fal/Video/MiniMax. The key goes in the .env next to run_nvidia_gpu.bat:

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

Where people get burned

A photoreal first frame gives you photoreal-ish low poly. The first frame anchors the whole clip. If you want the era's look, feed it era-appropriate art or feed it nothing.

Fast camera moves expose the geometry. Low-poly models simplify form; a whip pan or a hard orbit shows you every shortcut. Slower, wider camera language reads much better in this style - and it's a prompt-level fix, not a settings one.

Assuming 15 seconds will stay coherent. A stylised endpoint hides a lot, but the longer the clip, the more a single action wants to become three. Two 7-second clips beat one 14-second one when the story has beats in it.

Fixed 768p. No higher tier exists on this endpoint. Upscale locally if you need it; low-poly footage takes upscaling unusually well because there's no fine texture detail to smear.

Nothing is free about iterating. Every attempt bills per generated second. Draft the motion at 5 seconds, then re-run the keeper at the length you actually need.