"Someone Wakes" - Music Video: Hard-Shot Test Plan

"Someone Wakes" - Music Video: Hard-Shot Test Plan

Status: working plan. Nothing set. (2026-10-08, after John: *"isolate the shots that will be difficult to create and test them before committing to the full production. If we can't get something to work we find another way to tell the story."*) Companion to: `music-video-concept.md`, `music-video-storyboard.md`.

---

Why this exists

A full-motion video is a big commitment. Before we shoot 3 minutes, we isolate the shots most likely to break and test them in isolation. Each test ends in a verdict: works / works-with-a-change / find another way. Nothing graduates to full production until its hard shot has a verdict.

The hard shots, ranked by risk

1. Identity across shots - the same Dakota, one face, holding across many frames with wardrobe changes. *If this fails, the film's spine fails.* HIGHEST RISK. 2. She vanishes in-frame - a departing Dakota stops, looks, and is gone - *not* a dissolve, and *not* an edit. Must read as "no longer held by the frame." 3. The relay in one move - leave frame on one Dakota, arrive on the next already working, in one unbroken camera move. No cut. 4. The room holds constant - the same room, same warm-lamp/cold-window, every cycle. 5. The look into the lens - a calm, held gaze at camera; the thesis shot. Must not read as "acting." 6. The empty-room ending - hold on the room after D leaves; the open door; light through it.

Test log

*(each test: what we ran, what we saw, verdict, fallback if it failed)*

T1 - Identity + wardrobe (same face, different life)

• **Goal:** one face holds across two wardrobe states in the same room.
• **Setup:** day image (`image.png`) as the room; the likeness/portrait references as the face. Generate A (soft/domestic) and D (unadorned/open) in the same room, same light.
• **Pass criteria:** a viewer would read *the same woman* in both, dressed differently.
• **Verdict:** _running_
• **Fallback if it fails:** lock identity by generating each shot as an **edit of a single master still** (same base image, wardrobe swapped) rather than fresh generations; or accept a **single-wardrobe** film and carry "different lives" by *bearing and blocking* instead of dress.

T2 - The vanish (stop → look → gone, no dissolve)

• **Goal:** a departure that reads as absence, not effect.
• **Approach to test:** shoot her standing still, then the *room without her* - and let the **transition** be light/occlusion (something crossing the lens, a shift of the lamp), not an opacity fade.
• **Verdict:** _pending_
• **Fallback:** if no tool gives a clean in-frame vanish, **imply it** - cut to the room empty, and let the *look* do the work; the audience fills the gap. (Arguably truer anyway.)

T3 - The relay move

• **Goal:** one unbroken move from leaving-frame to arriving-on-the-next.
• **Approach to test:** a single camera move (pan/dolly) across the room; the next Dakota already present at the far end. Test whether a text-to-video / image-to-video tool can hold the room stable across the move.
• **Verdict:** _pending_
• **Fallback:** if the move can't hold the room, **split it** - end on the empty frame, begin on the next - which is a cut, but a *motivated* one (the room is the constant; the cut lands on it).

T4 - The room constant

• **Goal:** recognisably the same room every cycle.
• **Approach to test:** lock one **master room still** and drive every shot as an edit/variation of it.
• **Verdict:** _pending_

T5 - The gaze

• **Goal:** a calm held look into the lens.
• **Approach to test:** image-to-video from a still where she faces camera, prompt *"she slowly turns and holds a steady gaze into the lens, still, unblinking, no expression change."*
• **Verdict:** _pending_

T6 - The ending

• **Goal:** the empty room, the open door, light through it.
• **Approach to test:** a slow push-in on the room with the door open, no figure.
• **Verdict:** _pending_

T7 - Local animation (Wan 2.2 i2v) - FIRST LOCAL MOVING IMAGE (2026-10-08, 20:55)

• **Goal:** does the local stack animate our own stills at zero cost, and does it hold the room/light?
• **Run:** plain room plate → `mlx-video` Wan 2.2 TI2V 5B (Turbo LoRA, 4 steps). 2 s, 1280x704, 49 frames. **Total 119 s** (denoise 16.7 s, VAE 76.6 s).
• **Verdict:** ✅ **works - the room holds, the light holds, the camera pushes in.** First moving image made on the free rail. Free; private.
• **Cost:** $0 (local). Contrast: cloud video is metered per clip.
• **Artifact:** `video/room/sw-room-anim-v1.mp4` (+ first frame).
• **Note for scale:** a full ~3-min film at this rate is ~90 clips x ~2 min = ~3 h of Mac time, free. Feasible as a blockout engine.
• **Gotchas (verified):** run the **venv python directly** (never `uv run` mid-job); detach with **`screen`** (no `setsid` on macOS); **one GPU job at a time** (an idle ACE server killed attempt 2 with a Metal timeout).

Tooling notes

• **Image:** OpenAI (`gpt-image-2`) configured; Google (`gemini-3.1-flash-image`) configured - both can edit from reference images (the identity/room refs).
• **Video (LOCAL, free - the storyboard rail): `mlx-video` Wan 2.2 TI2V 5B** at `~/video-wan/mlx-video-gen/` (`generate.py`, `--image` = i2v). Proven 2026-10-08.
• **Video (cloud, paid - finals only): Google Veo 3.1** configured (i2v, up to 8 s). Others unset.
• **Images (LOCAL, free): mflux** - Qwen-2.1 (identity anchor), FLUX.2 klein Edit (restage), Krea 2, Qwen-Edit. Plus trained Dakota LoRAs in `~/lora/`.
• **Images (cloud, paid - finals only): google/gemini-3.1-flash-image**, openai/gpt-image-2. Ledgered by `scripts/gen-cost-ledger.py`.
• **Reference assets:** face = `projects/bands/drive-thru-outlaws/images/Dakota/Dakota_02_Closeup_Portrait.JPG` and the turnaround; room = `days/2026-10-07-someone-wakes/image.png`; wardrobe = `reference-packs/reference-pack-dakota-wardrobe.png`.

Provenance

Created 2026-10-08 in response to John's method. Nothing locked.