arkennemasis fal · VEED Fabric 1.0 Text · $0.08-0.15/s

This is the version of Fabric you actually want if you're building a talking head from scratch. The other two nodes in the family need an audio track - which means a TTS node upstream, an extra dependency, an extra failure point, and an extra set of takes to manage. Here you paste the script into a text box and the model writes the voice itself.

Image in, text in, talking video out. It's the shortest path from "I have a portrait and a paragraph" to "I have a video of that person reading the paragraph," and at 480p it's $0.08 per second of generated speech.

The inputs

image is the person, and it's required. There's no prompt to rescue a bad photo, so this is where the quality is won or lost: front-on, evenly lit, mouth visible. Feed the model a profile shot or a heavily filtered selfie and you'll get exactly what you paid for.

text is the script. Plain speech - the way you'd want a voice-over read, with punctuation doing the pacing work. The output length follows the rendered speech, and so does the bill, so a wordy paragraph is a longer, more expensive clip.

voice_description is optional, and this is the clever bit: the primary voice description is auto-generated from the image, and this field adds to it. The tooltip's own examples span simple descriptors like "British accent" or "Confident" all the way up to a full description ("Confident male voice, mid-20s, with notes of…"). Leave it empty and you take whatever the model infers from the picture - consistent, but not chosen.

resolution is 720p (default) or 480p. That's the entire settings surface. No prompt weighting, no seed, no motion strength. The model is doing the acting; you're choosing the face, the words and the voice flavour.

Outputs: video, video_path, and info, the JSON blob holding the request id, the cost estimate, the saved files and fal's full answer. Files land in output/fal/veed_fabric-1.0_text/.

Plus the pack-wide controls: max_cost_usd (default $20 cap - worth lowering while you experiment), reuse_identical_run, extra_json, max_concurrent.

When this loses to the audio-input version

If the voice matters - a specific person, a specific performance, a real recording - you want the plain Fabric 1.0 node and an audio track. Fabric's own launch pitch was against Kling Speak and the local avatar crowd on delivering a talking video, and the voice you get here is generated, not cloned. For anything with a brand voice or a real narrator, generate the speech separately (there's a load of options: local Chatterbox if you want it offline, or any of the pack's hosted TTS endpoints) and feed the audio in.

For placeholder narration, storyboards, internal demo videos, and the twenty-avatars-a-day case, the text node is simply faster to live with - one node instead of three.

Install

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

Manager → search arkennemasis. Then:

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

in the .env next to run_nvidia_gpu.bat, restart ComfyUI, and look under arkennemasis/fal/Lip Sync/VEED.

Notes from using the family

• Cost tracks the rendered speech, not your word count - so trimming a rambling sentence saves real money at $0.15/s on 720p. Set max_cost_usd before you paste in a five-hundred-word monologue.
• The voice_description field only adds to an auto-generated description. It won't replace the voice the image implies; it colours it.
• Everything is cached by default. reuse_identical_run returns the files from an identical earlier run instead of paying again - including after a ComfyUI restart. Editing the text or the image is what forces a new call.
• Long renders survive a crash on fal's side. The request id is logged in output/fal/_requests.jsonl; fal Recover Result downloads the finished job for free.