API

4c. Actor swap: your character's face on real footage

Mode v2v with "edit": "face" and product h3_swap: the largest face on every frame of your video is replaced by the character's face, built from its photos (or from ref_images, 1 to 8 clear photos), then enhanced. No diffusion: the footage, the body, the clothes, the camera and the soundtrack are exactly the original. This is the deepfake technique, and the reliable way to "change the actor". Price is per second and cheap (a few minutes of GPU). Add "voice": "character" to re-voice the speech too.

{ "mode": "v2v", "edit": "face", "product": "h3_swap", "seconds": 10, "reference_url": "https://aimodels.cheap/in/12-….mp4", "character": "chr_…", "prompt": "actor swap" }

Face lock. On text, photo and reference jobs with a character, "face_lock": true (the default) runs the same swap on the generated clip: the identity stays pixel-stable from clip to clip, on top of the trained adapter. Send false to keep the raw generation.

4d. Edit mode (experimental): regenerate the person or the background

Not recommended. In our tests the regenerated region comes out dark and incoherent, so the studio no longer offers it. To replace the whole actor use reference mode (4b) with the video plus character photos: the shot is regenerated with your character, body, shadows and interactions included. To keep the footage untouched and change only the face, use the actor swap (4c). The API still accepts the fields below.

Mode v2v keeps your video pixel for pixel and regenerates only one region: the main person ("edit": "person") or everything except the people ("edit": "background"). The person is segmented automatically on every frame; the H3 control network paints the new content inside that mask with the original footage as context, then the result is composited back onto your original frames. Camera, timing, lighting and the untouched areas are exactly yours. Add up to 5 photos of the new person (or of the new place) so the model copies an identity instead of inventing one; the prompt describes what goes inside the mask.

{
  "mode": "v2v",
  "product": "h3_edit",
  "seconds": 8,
  "reference_url": "https://aimodels.cheap/in/12-….mp4",
  "edit": "person",
  "ref_images": ["https://aimodels.cheap/in/12-a.jpg", "https://aimodels.cheap/in/12-b.jpg"],
  "keep_audio": true,
  "prompt": "same outfit as in the video"
}

Options: target = main (default, the largest person) or all (every person); keep_audio keeps your original soundtrack (default) or uses the generated one; control_strength 0.2 to 2 (default 1) trades faithfulness to the source for freedom. The output length follows the video (15 s max); seconds caps it. Aspect is taken from the video.

4e. Characters: train your own person once, reuse it everywhere

A character is a small adapter (LoRA) trained on 20 to 100 photos of one person, on a spot GPU, in about an hour. Once ready, pass "character": "chr_…" on any job: the adapter is stacked on the model and its trigger word is prepended to your prompt, so the same face, hair and body come back in text-to-video, image-to-video, reference and edit modes without sending photos again. Photos: varied angles, lighting, distances, expressions; sharp faces; no other people in frame. Upload them with /v1/uploads first. Each photo may carry a caption saying what it shows: outfit, pose, place, or the body part in a close-up. Captions keep the outfit and the setting out of the identity, so vary them freely; photos without a caption are described automatically on the GPU. Close-ups of hands, face, eyes, mouth, marks or any body part improve their consistency, keep them to about a third of the set and caption them as close-ups.

POST /v1/characters
{ "name": "Léo", "description": "young man, dark curly hair, slim",
  "images": ["https://aimodels.cheap/in/12-a.jpg", { "url": "https://aimodels.cheap/in/12-b.jpg", "caption": "close-up of his left hand holding a cup, outdoors" }, "…"] }

→ { "character": { "id": "chr_…", "status": "training", … }, "job": { … } }

GET /v1/characters            list (status: training | ready | failed)
DELETE /v1/characters/:id     remove the adapter

POST /v1/jobs  { "prompt": "walks along a beach at dusk, handheld", "character": "chr_…", "character_strength": 1.0, … }

To improve a character with more photos, send the full photo set (old and new) with "continue_from": "chr_…": training starts from that adapter with 200 steps and a gentler learning rate, keeps the trigger word, and produces a new character. Price: a flat fee per training, product h3_train in /v1/prices (about 1.5 GPU-hours at spot cost). Refunded if the training fails. The adapter is private to your account and never served publicly.

4f. Voice: the source video's speech, in your character's voice

Give a character a voice sample, 5 to 25 seconds of clean speech: PUT /v1/characters/:id/voice with the audio (or video) bytes and its content-type. Then, on reference (r2v) and edit (v2v) jobs that have a source video, "voice" chooses the result's speech: character converts the source speech into the character's voice, keeping the words, timing and intonation (zero-shot voice conversion, English works best); original keeps the source track; generated lets the model produce its own soundtrack. The chosen speech is also fed to the model as an audio reference so the mouth follows it. Default: character when the selected character has a voice sample, else original for edits and generated for reference jobs.