| name | venice-studio |
| description | Create visual and audio media with the Venice AI API — generate an image, place a consistent character into new scenes, turn stills into video (image-to-video), record a voiceover (TTS), compose background music, and assemble it all into a finished video, ad, or storyboard. Use whenever the user wants to make an image, a character image, a storyboard, a video, an ad/trailer, a voiceover, or a music bed. |
Venice Studio
One key, the whole pipeline: images · character compositing · image→video · voiceover · music · assembly. Everything routes through two scripts in this skill's folder. Stdlib Python + ffmpeg, no install.
SKILL_DIR=~/.claude/skills/venice-studio
python "$SKILL_DIR/venice.py" <image|edit|video|tts|music|models|voices|balance> ...
python "$SKILL_DIR/assemble.py" <kenburns|clip|concat|vo|loopfit|mix|caption|probe> ...
The key auto-resolves from $VENICE_API_KEY or a VENICE_API_KEY= line in a .env (cwd or any parent). Each Venice command prints the output file path; run venice.py balance to see USD left.
When the user asks for…
- an image →
venice.py image --prompt "…" --out x.png (default model flux-2-pro; try nano-banana-pro, ideogram-v4, seedream-v4).
- a specific character in a scene / keep a character consistent → start from a reference PNG and
venice.py edit --image ref.png --prompt "place this exact character …, keep its design identical" --out shot.png. nano-banana-pro-edit is excellent at holding a character across shots — this is how you get the SAME character in every frame.
- a voiceover →
venice.py tts --text "…" --voice Brian --out vo.mp3 (run venice.py voices for the list; Brian/Bill/George/Chris are strong; avoid kokoro — it sounds robotic).
- music →
venice.py music --prompt "…" --duration 20 --instrumental --out bed.mp3.
- a video / ad / trailer / storyboard-to-video → follow the workflow below.
Workflow for a full video (image → motion → voice → music → cut)
- Brief. Pin down: style/tone, length, the character(s), whether there's a voiceover (and voice), and music mood. Ask only what you can't reasonably choose.
- Storyboard. Write the shot list — for each shot: the visual, any motion, the VO line, any on-screen text. Share it before spending.
- Stills (cheap, ~$0.04–0.20 each). Generate every key frame:
image for new scenes, edit (with a reference) for any shot featuring the character. Show them to the user before the expensive step.
- Motion (costs real money — ~$0.76 per 5s clip; confirm first). For shots that should move:
venice.py video --image still.png --prompt "<the motion>" --duration 5s --out clip.mp4. For static shots, the cold open, and anything with a human face (Venice gates face→video behind a consent attestation), skip generation and use a Ken Burns push-in instead: assemble.py kenburns --image still.png --duration 6 --out clip.mp4 [--bw].
- Voiceover. One
tts call per line or per storyboard beat (separate files make timing easy).
- Music. One or two
music beds (e.g. an energetic bed + a soft bed for a tonal shift).
- Assemble:
- normalize each shot to a clip:
kenburns (stills) or clip (trim/normalize a generated mp4),
concat --inputs s1.mp4 s2.mp4 … → one silent video,vo --inputs l1.mp3 l2.mp3 … → one voice track (auto-gaps),loopfit each music bed to length (with fades), thenmix --video silent.mp4 --vo vo.mp3 --music bed.mp3 --out cut.mp4 (music auto-ducks under the voice),caption --in cut.mp4 --text "…" --start S --end E for on-screen text/end cards.
- Review & iterate. You can't hear audio or watch motion yourself — extract a frame (
ffmpeg -ss T -i v.mp4 -frames:v 1 f.png) to spot-check visuals, then hand the file to the user and ask what to tune (voice, pacing, a reshot frame).
Money & manners
- Images and TTS are cents; video and music cost real money. Before generating a batch of clips, state the rough total and get a go. Check
venice.py balance first and after.
- Put working files in a dedicated folder (a project
media/ dir or the session scratchpad), not scattered in the repo.
- Never wildcard-delete near your output (
rm s*.mp4 once ate the final cut — its name started with s). Remove intermediates by explicit name.
Catalog (as of 2026)
32 image · 100+ video (seedance, wan — image-to-video & reference-to-video) · 11 music (elevenlabs-music, lyria-3-pro, stable-audio-25) · 11 TTS voices. List any with venice.py models --type <image|video|tts|music>.
Worked example: the Starchild "PurposeMaxxing" teleshopping ad was built entirely this way (~$5 total) — see the venice-media-ad-pipeline memory for the concept and per-step costs.
