TouchDesigner component

TD Fal

1,400+ fal.ai models in one TouchDesigner component.

Load a schema, pick a model, hit Generate. TD Fal reads each model's input schema and builds real TouchDesigner parameters for it, so Gemini images, Kling video, Hunyuan3D, ElevenLabs voices and Flux all work from the same component.

Get it on Gumroad Features
td-fal.toe — /project1
The TD_Fal component in TouchDesigner with its Connection page, generated image and video outputs, and the activity log
1,400+
fal.ai models
6
modalities: image, video, 3D, vision, speech, audio
8
parameter pages
0
request bodies to write

Features

  • 01

    Schemas become parameters

    Every fal endpoint declares its own inputs. TD Fal fetches the selected model's schema and generates real TD parameters for every field: enums as menus, numbers with their ranges, booleans as toggles.

  • 02

    Future-proof by design

    Nothing is hard-coded per model, so models released after the component was written just work, and every field of every model is reachable.

  • 03

    Drive it like TouchDesigner

    Because they're real parameters, they take expressions, exports and CHOPs like anything else in your network. Animate a seed. Map a knob to guidance.

  • 04

    Image, video, 3D, voice, audio

    Outputs land as TOPs, CHOPs, glTF and DATs you can wire straight into your network. The Any Model page takes a bare endpoint id plus raw JSON for everything else.

  • 05

    Live model filter

    Type flux, kling or elevenlabs and every dropdown refilters instantly from the cached index. No re-fetch.

  • 06

    Never stalls

    Jobs run on background threads and report back through a queue drained on frame start. Poll interval, max wait, queue priority and abort-all are all on the Connection page.

Outputs

Drop it in. Wire it up.

Every output is a regular TouchDesigner operator, ready to wire into the rest of your network.

BaseTD_FalimageOutvideoOutimage_schemavideo_schemalogprogress

Why it's built differently

Most AI APIs share one request shape across all their models. fal doesn't: a Kling video model wants start_image_url and a string duration of "5", Hunyuan3D wants input_image_url, ElevenLabs wants text instead of prompt.

So TD Fal doesn't hard-code request bodies. When you generate, it:

  1. fetches the selected model's OpenAPI schema from fal,
  2. works out which of that model's fields play the role of prompt, reference image, seed, duration or voice,
  3. generates real TouchDesigner parameters for every other field the model declares,
  4. builds the request from all of the above, with your Extra JSON merged last.

Parameter pages

PageModels
Imagetext-to-image and image-to-image (580)
Videotext, image, video and audio to video (530)
3Dimage-to-3D, text-to-3D, 3D-to-3D (53)
Visionvision, image/video to text, LLM, JSON (56)
SpeechTTS, speech-to-text, speech-to-speech (48)
Audiomusic, SFX, stems, audio-to-audio (95)
Any Modelany endpoint id plus raw JSON, including training and workflows