Veed Fabric: how to write prompts the model actually understands

How Fabric works and how it differs

Fabric is not text-to-video. It is a lip-sync / talking head system built on a Diffusion Transformer (DiT) architecture that animates a still image to audio. The input is a pair: one image and one audio file. The model produces lip-sync, adds natural head and body motion, and adds hand gestures tied to the speech rhythm.

The image can be in any style: photo, illustration, anime, 3D render, clay mascot, brand character. This is the key difference from classic avatar generators — Fabric does not require a photorealistic face. Audio is speech or music. Resolutions are 480p and 720p, frame rate 25 fps, formats 16:9, 9:16, and 1:1. Speed: Fabric 1.0 Fast is roughly 2.5× faster than Standard, the 480p path handles 10 seconds of video in about 1.5 minutes.

• Image + Audio → lip-synced video (not T2V)
• Any input image style — photo, illustration, anime, 3D
• 30+ languages, up to 5 minutes via API
• Resolutions 480p and 720p, 25 fps, formats 16:9 / 9:16 / 1:1
• Fast variant runs ~2.5× faster than Standard

Two working modes

Audio mode: you bring your own audio file (mp3/wav/m4a/aac/ogg, up to 10 MB). There is no text prompt here at all — the model only syncs lips and motion to the sound. Quality depends on the image and audio pair: clean recording with no background noise, frontal image with a visible face, no extreme angles.

TTS mode (via VEED): audio is generated from a speech script by the ElevenLabs V3 engine. Here the «prompt» is the script itself: the text spoken on screen plus inline bracketed tags that control emotions, pacing, accent, and sound effects. The script can be in any of the 30+ supported languages.

Fabric Emotions: inline tags inside the script

In TTS mode the script carries [tag] markers for emotional expression. These are not formatting markers — they are directorial cues for the voice engine:

Emotions: [excited], [happy], [sad], [angry], [curious], [nervous], [confident]. Reactions: [laughs], [sighs], [gasps], [clears throat]. Volume: [whispers], [shouting]. Pacing: [pause], [long pause], [breathes], [rushed], [drawn out]. Sound effects: [applause], [gunshot], [door creaks]. Accent: [American accent], [British accent].

One rule: do not overload. One tag per 1-2 sentences, distributed gradually for natural delivery. A tag before every word will break the intonation — the model will start clipping each word with awkward pauses and gear shifts.

Requirements for the input image

The image sets the entire visual quality. Base rules: frontal or near-frontal face, no heavy Dutch angles, no 90-degree profiles. Face well lit — no deep shadow on one half. No occlusions (hands in front of the face, masks, glasses with strong reflections on the eyes) — the model cannot pull lip-sync if it cannot see the mouth.

Formats: jpg, jpeg, png, webp, gif, avif, up to 10 MB. Style is not critical: Fabric animates photos, anime illustrations, 3D-rendered clay characters, and corporate mascots equally well. But in every case you need one clear face in frame — not a crowd, not two characters, not a profile without a visible mouth.

Common mistakes

1. 1. Describing a scene instead of writing a script

   Fabric is not T2V. A prompt like «a man in a forest at sunset, walking and explaining the product» will be ignored: the model does not generate the forest, the sunset, or the walking. Hand in a ready image (background plus face) and a speech script with tags. Some other tool draws the scene; Fabric only animates the face.

2. 2. Overloading with emotion tags

   [excited] Hello [happy] everyone [laughs] today [confident] I want — the model will jitter, pause, and break into unnatural transitions. One tag per 1-2 sentences, not per word. Place reactions like [laughs] and [sighs] between phrases, not inside them. Tags work as director's notes, not as per-token markup.

3. 3. Dirty audio or heavy background noise

   In Audio mode lip-sync quality depends directly on how clean the sound is. Heavy background noise, echo, or music layered over speech confuses the model — lips start to drift and sync breaks. Record speech alone, add background music in post after generation, not into the source audio file.

4. 4. Extreme angles in the input image

   A strong profile, Dutch angle, or face occluded behind a hand, mask, or reflective glasses — the model cannot pull lip-sync. Use a frontal or near-frontal image with a clearly visible mouth and even lighting. Art style is not critical, the angle is.

5. 5. Expecting camera moves or action

   Fabric does not do dolly, push-in, or tracking shots — the model does not move the camera and does not change the shot. If the brief needs cinematic motion and physical action in an environment, that is a job for Sora 2, Veo 3.1, or Kling. Fabric covers a different case: a fixed frame brought to life by face and speech.

Before / after examples

person talking about a product

[TTS script for Veed Fabric, paired with a frontal product-shot image of a brand mascot]

[confident] Hey there! I'm Otto, and today I'm showing you something special. [pause] Our new wireless earbuds give you twelve hours of battery on a single charge. [excited] Twelve full hours — that's almost a whole workday! [pause] [drawn out] No more low-battery anxiety. Tap the link below to grab yours.

a brand mascot says hello to viewers

[TTS script paired with a frontal illustration of the brand mascot]

[happy] Hello, friends! [laughs] It's so good to see you again. [pause] I've been waiting all week to share this with you. [curious] Have you ever wondered what makes our community special? [pause] [confident] Stick around — I'll show you in the next sixty seconds.

explain something in two languages

[TTS script for Veed Fabric, paired with a frontal image of an animated instructor — illustration style]

[British accent] [confident] Welcome back to the channel. Today we're tackling something most beginners get wrong. [pause] [curious] What if I told you the trick is in the timing, not the tools? [drawn out] Let me show you. [pause] In the next clip I'll walk through it step by step.