An educational framework
The Tiny Desk AI Prompt, Shot by Shot
Describe who belongs in each role, which parts of the reference should guide the edit and what must remain consistent. A useful prompt separates those decisions. It is not a magic sentence that guarantees matching faces, clean hands or the right music.
On this page
One reference edit is different from five independent generations
This site submits a complete reference-based edit for the selected cast. It does not generate five independent shots and stitch them together as its standard workflow. Audio is handled after the visual render.
The blocks below are a learning framework for planning your own experiments. They are not this site’s complete production prompt, hidden model settings or a copy-and-run recipe. A different model may need different inputs or may not support a video reference at all.
1. Define the people and their roles
Give each person one stable label and use that label consistently. A role map should say both which photo identifies a person and where that person belongs in the performance. Do not rely on the order of an unnamed attachment list.
Keep identity descriptions factual: visible hair, facial features and clothing. If you ask for a new outfit, make that a separate instruction so an identity change is not mistaken for a styling request.
CAST MAP
Lead: [portrait assigned to the lead singer]
Saxophone: [portrait assigned to the saxophonist, if supplied]
Keyboard: [portrait assigned to the keyboardist, if supplied]
Backing vocals: [left and right assignments, if supplied]
IDENTITY GOAL
Preserve each assigned person consistently across close and wide views.
Do not swap one assigned person into a different role.2. Describe what the reference controls
Break the reference into observable constraints: camera framing, sequence of cuts, timing, room layout and performer movement. Then state what the new inputs should change. An instruction such as “make it exactly the same” leaves the identity-versus-motion boundary unclear.
REFERENCE GOAL
Use the permitted reference for the performance timing and camera sequence.
Apply the cast map to the selected roles.
Keep the desk layout, instruments and lighting coherent across cuts.
REVIEW GOAL
Check faces, role placement, hands and instruments in every view.3. Review by shot, even when you generate the whole clip
If you use a separate-shot workflow in another tool, carry the same cast and scene definitions into every segment. Join clips only after checking continuity. A multi-shot strategy is an option, not a universal requirement and not a guarantee of better results.
| View | Ask yourself |
|---|---|
| Opening lead view | Does the lead match the assigned portrait without changing during motion? |
| Band view | Are the chosen people still in the intended roles? |
| Expressive lead close-up | Do the face and mouth stay plausible through the expression? |
| Saxophone focus | Does the right person appear, and does the instrument remain coherent? |
| Closing band view | Do the cast, clothing and positions agree with the earlier wide shot? |
4. Plan audio separately
A text prompt cannot grant permission to use a recording. Decide which audio you can use, whether the model preserves it and whether you need an editing step afterward. Check the completed file rather than assuming sound survives a visual edit.
In this site’s workflow, the reference audio is combined with the visual result during finalization. A pending audio or storage step is not yet a finished download. A separate silent export is not promised.
5. Change one thing at a time
If a face is unstable, first improve the portrait or simplify the cast instructions. If the wrong person appears in a role, fix the mapping before adding more stylistic detail. If the camera sequence changes, revisit which input supplies motion and timing.
Keep notes on the model, input photos and the change you made. This makes the next attempt a useful comparison instead of a completely different experiment. Do not assume a prompt that worked once will produce an identical result on another run.
The step-by-step checklist
- 01
Map the cast
Give each reference portrait a stable label and assign it to the intended lead, instrument or backing-vocal role.
- 02
Define the reference constraints
Identify which permitted reference controls framing, cuts, timing and motion, and which selected people the edit should change.
- 03
Review every shot
Check the same faces, roles, clothing and instruments in close and wide views, whether you generated the entire clip or separate segments.
- 04
Plan and check audio
Use audio you have permission to use, account for a separate finalization step and verify sync in the completed file.
- 05
Refine one variable
Improve the portrait, role map or motion instruction one at a time and compare results rather than changing every input together.
Questions, answered
- Is this the exact prompt used by the generator?
- No. This is an educational framework. It does not disclose the complete production prompt, provider settings or orchestration used by this site.
- Do I need five separate prompts?
- Not necessarily. Some workflows generate separate shots; this site uses a complete reference-based edit. Review every shot regardless of how the video was generated.
- Can the prompt guarantee the original song?
- No. Audio handling and permission to use a recording are separate from the visual prompt. Check the audio workflow and the finished file.
- Why does identity drift between shots?
- Ambiguous role assignments, weak reference portraits and model limitations can all contribute. Use clear inputs and consistent labels, and inspect the complete result.