> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tavus.io/llms.txt
> Use this file to discover all available pages before exploring further.
# Choose a Model and Training Path
> Pick a Phoenix model, then decide whether to train from a photo or a video.
Create a face with two choices, in order: which **Phoenix model**, then whether you start from a **photo** or a **video**.
Use the [Create Face](/api-reference/faces/create-face) API with `model_name` plus either `train_video_url` or `train_image_url` (never both). The [PAL Maker](https://maker.tavus.io/dev/faces/create) asks the same questions with live validation.
## Step 1. Pick a model
Newest model. Start using your face in minutes.
**Choose this if**
* You want a *zero-shot preview* you can use immediately, followed by a tuned version
* You want conventional 'talking head' framing: chest-up or waist-up, not head to toe
* You want movement below the neck (torso, shoulders, hair, & clothing) that is more natural
* You only have a photo to work from
**Do not choose this if**
* You need a full-body face, head to toe (that is Phoenix-4)
* You need clothing with complex texture to stay exact, like a small logo or lots of decorations (that is Phoenix-4)
* You cannot ship anything with a watermark on it, even briefly
The full-body option.
**Choose this if**
* You need full body or a wider body shot
* You need clothing with complex texture to stay as it is, like a small logo or lots of decorations
**Do not choose this if**
* You want to start using the face right away (Phoenix-4 is only usable when the whole training job finishes)
* Your photo or video footage has glasses, jewelry, headphones, hair over the shoulders, or a heavy beard (Phoenix-4 rejects those; Phoenix-4.5 does not)
Available for edge cases. Almost always use 4.5 or 4 instead.
**Choose this if**
* You have a specific reason to pin this older model. Those cases are rare.
**Do not choose this if**
* You want to use the latest, most expressive, and most realistic model
* You want to train from a photo (Phoenix-3 only takes video)
All models accept a **human or clearly human-shaped** subject: realistic, cartoon, anime, and Pixar-style faces are fine. Animals, mascots, and other non-human figures are not.
## Step 2. Pick an input for that model
Phoenix-4.5
A photo is faster to set up. Video takes more work, and the finished face usually moves more naturally.
Easier. One photo, no recording.
**`train_image_url`**. You also attach a voice. See [Voices for Image-Based Faces](/sections/faces/voices-for-image-based-faces).
* Preview usable in about a minute, then tunes in the background
* Chest-up or waist-up, not full body
Better result. Your real movement, and your own voice.
**`train_video_url`**
* Preview takes a few minutes, then tunes in the background
* Voice is created from the recording
* Chest-up or waist-up, not full body
Phoenix-4
Upload one photo. Ready when training finishes.
**`train_image_url`**. You also attach a voice. See [Voices for Image-Based Faces](/sections/faces/voices-for-image-based-faces).
* No early usable version and no watermark
* Stricter photo checks: hair behind the shoulders, neck visible, sleeves or wide straps, no glasses, no necklace, no large earrings, no over-ear headphones, no heavy beard covering the neck
Upload footage. Highest fidelity today, and the only full-body path.
**`train_video_url`**
* Thirty seconds speaking, then thirty seconds listening
* Full body: record vertically, head to toe, standing still
* A voice is created from the video's audio
* Ready when training finishes
Phoenix-3 (legacy)
Not available on Phoenix-3.
Training from a photo was introduced on Phoenix-4. If you want to start from a photo, use Phoenix-4.5 or Phoenix-4.
Video only. Legacy.
**`model_name`: `phoenix-3`**
* Ready when training finishes
* Older recording rules apply
* Almost always use Phoenix-4.5 or Phoenix-4 instead
## Complete face-creation procedure
This section is self-contained. Follow it to create a face correctly without loading another page. Page links are for extra capture detail only.
### Request shape
One call: `POST /v2/faces`. Send `model_name` plus **exactly one** of `train_image_url` or `train_video_url`. Never both. Both URLs must be publicly downloadable (for example a presigned S3 GET URL) and stay valid at least 24 hours after submission.
### Step 1. Pick a model
Stop at the first match.
1. Face must show a full body, head to toe → `phoenix-4` with `train_video_url`. This is the only full-body path. Record vertically.
2. Clothing has complex texture that must stay exact, such as a small logo or heavy decoration → `phoenix-4`. It copies the body region more faithfully.
3. A watermark is unacceptable even briefly → `phoenix-4`. Every `phoenix-4.5` face is watermarked during its preview stage.
4. Anything else → `phoenix-4.5`.
`phoenix-3` is legacy and video only. Use it only when explicitly asked to pin it.
### Step 2. Pick an input
* Only a photo available → `train_image_url`. Fastest path. On `phoenix-4.5` the face is usable in about a minute.
* Person can record footage → `train_video_url`. Slower to become usable, but the tuned face moves more naturally and a voice is created from the recording's audio.
Image training is not available on `phoenix-3`.
### Voice rules
* Image training has no audio to clone, so a voice is **required** on the same request: send `voice_name` (stock slug, for example `anna`) **or** `default_voice_id` (a Tavus voice ID). Send one, not both. `voice_id` is a PAL TTS field, not a Create Face field.
* Video training creates a voice from the video's own audio and sets it as the face's `default_voice_id`. Do not send a voice.
* Detail: [Voices for Image-Based Faces](/sections/faces/voices-for-image-based-faces).
### Phoenix-4.5 lifecycle and polling
One call produces two stages and the face ID never changes.
1. Zero-shot preview: usable early (about a minute from a photo, minutes from video). Watermark burned into the video.
2. Background tuning: a few hours. The unwatermarked version replaces the preview. No second API call.
Poll `GET /v2/faces/{face_id}`:
* `status` `started` → not usable yet.
* `status` `completed` → the preview is usable. This does **not** mean tuning finished.
* `finetune_status` `started` → still watermarked.
* `finetune_status` `completed` → unwatermarked version ready.
* `finetune_status` `errored` → tuning failed; the preview stays usable with a "tuning failed" watermark.
* `status` `error` → creation failed and the face is unusable; read `error_message`.
`finetune_status` is unused on `phoenix-4` and `phoenix-3`; those are ready only when `status` is `completed` (Phoenix-4 typically 3 to 4 hours). Detail: [Phoenix-4.5 Preview and Status](/sections/faces/phoenix-45-preview-and-status).
### Hard capture constraints
Apply these before submitting. They cause most rejections.
* Phoenix-4.5 framing: chest-up is best, waist-up or slightly closer is fine, face fully visible. Never head to toe.
* Phoenix-4.5 accessories: glasses, jewelry, headphones, and hair in front are allowed if they do not cover the face. Shorter, crisp beards that leave the mouth and teeth visible work best - results may vary with dense or long beards.
* Phoenix-4 is stricter: no glasses, no over-ear headphones, no large jewelry, hair behind the shoulders, neck visible, sleeves or wide straps, no dense beard over the neck. A photo that fails Phoenix-4 often passes Phoenix-4.5.
* Both models accept a human or clearly human-shaped subject: realistic, cartoon, anime, and Pixar-style faces are fine. Animals, mascots, and other non-human figures are not.
* Both models reject: multiple people, subjects under 18, obvious public figures or copyrighted characters, blurry or poorly lit images, off-center framing, and unnatural poses.
* Video, Phoenix-4.5 and Phoenix-4: one continuous take, 30 seconds speaking then 30 seconds listening, teeth visible while speaking, lips closed while listening, static background, posture and framing unchanged between segments. Phoenix-3 uses 1 minute speaking plus 1 minute listening.
* Optional on image training: `auto_fix_training_image` `true` adjusts a photo toward the rules. It cannot fix policy failures (under 18, copyright, non-human) and it will not invent a full body.
Per-model capture detail: [Phoenix-4.5 image](/sections/faces/phoenix-45-image-requirements) · [Phoenix-4.5 video](/sections/faces/phoenix-45-video-requirements) · [Phoenix-4 image](/sections/faces/phoenix-4-image-requirements) · [Phoenix-4 video](/sections/faces/phoenix-4-video-requirements) · [Phoenix-3 video](/sections/faces/phoenix-3-video-requirements)
### Minimal correct calls
```json Phoenix-4.5 from a photo theme={null}
{
"face_name": "my_image_face",
"model_name": "phoenix-4.5",
"train_image_url": "https://example.com/headshot.png",
"voice_name": "anna"
}
```
```json Phoenix-4 full body from video theme={null}
{
"face_name": "my_full_body_face",
"model_name": "phoenix-4",
"train_video_url": "https://example.com/training-video.mp4"
}
```
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.