ClipForge (under Produce → ClipForge) is two tools in one page:
  • Generate — AI video clips from a text prompt (text-to-video) or by animating one of your images (image-to-video).
  • Discover — one search box across Pexels, Pixabay, and Coverr free stock, with one-click import into your Library.
Gating: generation needs Pro and up (clipforge entitlement) and spends credits. Discover (stock search + import) is open to every member on every plan — no entitlement, no credits.

​
Generate — t2v and i2v

Text to video

Describe the clip — the starter chips help you build a strong prompt from shot, camera, style, and light vocabulary.

Animate an image

Pick an image from your Library and describe the motion; ClipForge brings the still to life (image-to-video).

​
Options

As with SonicBliss, the prompt library gives you operator starters, your saved prompts, and favorites; the dice drops in a random starter. Hit Generate and follow the job live in the Activity dock — each variation appears in the clips grid as it finishes. Animate an image (i2v): click Animate an image and the media picker opens — search your Library images by name, filter by project, sort, and preview any image full-screen before choosing. Pick one and describe the motion; the composer switches to image-to-video. Use Change to swap the image or Clear to return to text-to-video.

​
Character replace (v2v)

Pick a source video and a reference character image together and the composer switches to character replace: the person in your clip is replaced by your reference character, keeping the original movement, timing and audio (which is why the lip-sync still lines up). There’s no mode switch — picking both files is the switch. Aspect and length come from the source clip, and the run is fixed to a single take.
You must confirm you have the right to use the likeness. Replace character stays disabled until you tick the confirmation, and the tick resets whenever you change either file — consent is per-person, not per-session. We record that you confirmed it against the job so there is an accountable trail; we do not verify it, and the responsibility for having permission is yours. If character replace isn’t offered at all, your operator has disabled the models.
What it genuinely can’t do. Worth knowing before you spend credits:
  • Body shape is not retargeted. Your character is mapped onto the performer’s existing movement, so a very different build — much taller, much heavier, a child against an adult — visibly stretches or squashes to fit. Match the frame and it holds up far better.
  • Identity drifts on long clips. Past roughly a minute the face gradually stops matching your reference. Cut to the shot you need first — cheaper and better looking.
  • Hands and occlusion are the weak spots. Fingers, and anything passing in front of the performer, are where artefacts show up first.
Cost. This model is billed in 16-frame “video seconds”, so the price follows the source clip’s frame count, not its running time — 60fps footage costs ~3.75× the same clip at 16fps. Re-export at 24–30fps and trim to the shot you need. The composer shows a live estimate (always an estimate) before you commit.

​
Dub — put new words in an existing clip (#190)

Pick a source video and an audio track together (no character image) and the composer becomes dub: the speaker’s mouth is re-animated to match your new audio — a translation, a re-recorded line, a fixed take. Everything else in the shot is untouched. There is no mode switch; picking the two files is the switch, and the run is fixed to a single take.
You must confirm you have the right to use the likeness — the same tick as character replace, and it resets whenever you change any of the files, including the audio. New words in a real person’s mouth is a new thing to consent to. We record the confirmation against the job; we do not verify it.
What it genuinely can’t do.
  • Only the mouth changes. Head movement, gestures and timing keep doing exactly what they did in your clip, so a new take that is much longer or shorter than the original still runs against the old performance. Record roughly to length.
  • The audio decides the length, and the price. The result covers the longer of your clip and your audio. Trim the audio to the line you need first.
  • Faces must be visible and roughly front-on. Profiles, motion blur, a hand or mic across the mouth, or several faces in shot are where artefacts appear.
Cost. Billed per second of output, not by frame count — frame rate and resolution make no difference here. The default model is open-source and roughly 10–16× cheaper than the proprietary tiers offered beside it; start there and only opt up on footage it struggles with. The composer shows a live estimate (always an estimate) before you commit.

​
Perform — make a photo deliver a recording (#191)

Pick a still image and an audio track (no video) and the composer becomes perform: the person in the photo delivers that audio, with expression and body movement driven by the sound. The output is exactly as long as the audio.
This is the most expensive model on the platform — roughly 0.16foreverysecondoffinishedvideo(about0.16 for every second of finished video (about 9.60 a minute). A one-minute clip costs about ten times a six-second one. Check the estimate, and cut the audio down before you run it. It is also likeness-gated, with the same non-verified attestation as the other lanes.
What it genuinely can’t do.
  • One person, clearly framed. Several faces, a very small face, or a heavily obscured one give the model nothing stable to animate. Head-and-shoulders or half-body portraits work best.
  • It invents the motion. Gestures, posture and expression are the model’s guess from what it hears, not a performance you directed — and hands stay the weakest part. Treat take one as a draft.
Type once, search Pexels + Pixabay + Coverr together. Filter by orientation and minimum duration; results are cached for 10 minutes, so refining a query is instant.
Query library & quick chips. Not sure what to search? Tap a quick-pick chip under the search bar to run a curated query instantly, or open the Query library for the full set of starter search queries (grouped by category) plus anything you’ve saved. A starter’s query and its orientation/duration filters apply in one tap. After any search, Save current search stores that query and its filters as your own preset.Resize the search field. Drag the grip on the right edge of the search box to make it as wide as you like (your size is remembered per browser; double-click the grip to reset). On mobile the field is always full-width.
1

Search

Enter a query (and optional orientation / min-duration filters). Results from all three providers appear in one grid, each labeled with its source.
2

Check the license

Every result carries its license and whether attribution is required. All three sources are free to use; Pexels requires attribution, Pixabay and Coverr don’t.
3

Import

One click imports the clip into your Library — with its license and source recorded as provenance, so the attribution answer travels with the asset forever.

​
Your clips

Generated and imported clips land in the Library as video assets and in the ClipForge grid:
  • Hover a card to preview; double-click (double-tap or the expand button on touch) opens a popup player.
  • Every clip shows a license/source badge, and generated clips carry full provenance (gen_meta): model, license, commercial-use status, prompt, and seed.
  • One click hands a clip off to the VEditor to build it into a cut.

​
Credits

Generation spends from a monthly credits budget (1 credit = $0.01): Video costs ≈8 credits per second generated (audio in SonicBliss ≈0.5/s); operators can tune the rate per model. Overage is covered by credit packs. Discover imports are free.

​
Troubleshooting

  • “Model busy” — the provider is under load; retry in a moment or switch models.
  • “No key” / provider not configured — the instance operator hasn’t configured that model’s provider yet. Pick another model, or contact support so an operator can enable it.
  • Generation locked but Discover works — that’s the plan gate: stock search/import is free for everyone, generation (and pricing a generation) needs Pro+.
  • “Replace character” is greyed out — tick the likeness confirmation under the character-replace controls. If you just swapped the source clip or the reference image, the tick was cleared on purpose and needs confirming again for the new person.

​
Over the API

The routes appear in the API Reference tab under the clipforge and genai tags:
  • POST /clipforge/generate → 202 {job_id} — mode (t2v|i2v|v2v|dub|perform), prompt, image_asset_id (for i2v, the reference character for v2v, the portrait for perform), video_asset_id (the source clip for v2v and dub), audio_asset_id (the new audio for dub, the performance for perform), aspect, duration_seconds (1–30), negative, model_key?, variations (1–4; forced to 1 for v2v/dub/perform), seed?. Any mode that alters a real person — v2v, dub and perform — also requires likeness_consent: true; without it you get a 422 with error: "likeness_consent_required". The confirmation is written to the audit trail against the job_id. dub and perform take no prompt and ignore aspect/duration_seconds: their length (and their bill) comes from the audio, and the model row’s max_seconds caps it.
  • POST /clipforge/quote — price a /clipforge/generate body before you send it. Same body, same model resolution, always an estimate. Needs the same Pro+ clipforge entitlement as generation.
  • GET /clipforge/clips — list generated clips.
  • GET /clipforge/stock/search?q=&orientation=&min_duration= — one search across Pexels/Pixabay/Coverr; results carry license + attribution_required; cached 10 min.
  • POST /clipforge/stock/import {source, id} → 202 — import with license provenance.
  • GET /genai/models?feature=&mode= · /genai/prompts (+ PUT /genai/prompts/prefs) — models and the prompt library.
Follow any job_id on GET /jobs/{id}/events (SSE) — see the Developer guide for a full recipe.