Case 02 · Concept · AI product film
HOVER
Lighter after three.
Format
0:40 · 16:9 · 4K · 30 fps
Mixed to −14 LUFS
Role
Concept, script, previs, keyframes, generation, edit, grade, sound
Tools
Blender, GPT Image 2, Seedance 2.0, ElevenLabs, Real-ESRGAN, After Effects
Brand
Invented for this film. A concept, not a client job
The 3 p.m. slump, played as a meeting.
HOVER is an invented sparkling cold brew, and the brief I gave myself was the afternoon slump everyone has sat through.
A boardroom where everyone has melted into their chairs and the steam pours down from the mugs instead of up. One person is still awake. She doesn’t argue — she slides a can across the table. One sip, and the slump turns into a rush you can see: the light, the tie, the papers, the work.
A slump, a slide, a rush, and a pass.

A · Meeting
Steam pours down. Only Maya is awake. A sigh, a look at Leo — and a can slides across the table.
0:00–0:12
B · Wake-up
One sip, one take: the light turns warm and the rush hits.
0:12–0:20
P · Inside the can
Cold brew, ice and coffee beans in zero gravity.
0:20–0:28
E · Next morning
Leo walks in with one can, spots Tom asleep and lobs it over.
0:28–0:40Every shot is blocked before it’s generated.
Each scene starts as a grey Blender blockout: the long boardroom table and who sits where, real lenses from 28 to 70 mm, every camera move, and the moments that matter timed to the frame — the sigh, the slide, the rush, the lob.
The blockout is the motion reference for the video model. Even the tired people move in it — rubbing eyes, nodding off, yawning — because a still blockout gives you statues.



Melted meeting



Maya presents



Nora yawns, Leo drifts



The sigh



She turns to Leo



A can from under the table



Slid to Leo



The can arrives



Still asleep



Ring-pull



Still drinking



The rush



Collecting himself



Into the work



Zero gravity



The orbit



Packshot



Next morning



He spots Tom



Over the shoulder



The lob
One can, drawn the same in every shot.
Maya, Leo, Tom, Nora, the boardroom and the office are designed once, as reference sheets. The can is one simple design — a big vertical wordmark and a coffee bean, nothing else printed — and that single image goes into every frame that shows it, so the label never drifts.
7 sheets
One can, four people and two rooms for all 23 keyframes and four scenes.
23 keyframes, approved before anything moved.
A frame for every camera change and every key state, made with GPT Image 2. Each frame is built from the previous one in the same room, so the furniture, the mugs and who sits where never jump between cuts. Cold overcast light until the sip, real warm sun after it. Click any frame to open it.
A
Meeting · 7 frames · GPT Image 2B
Wake-up · 8 frames · GPT Image 2P
Inside the can · 3 frames · GPT Image 2E
Next morning · 5 frames · GPT Image 2Prompts written like a shot list.
One generation per scene, several shots cut inside a single take, each with its camera move and an end state. The storyboard frames and the can go in as references — no character portraits, which the model likes to paste in as a still. Physics and pacing are written out: real-time speed, one can, the hand travels with the can.
Same low front-left angle, waist-up, a moment later: Leo (man, 32, short tousled brown hair, light stubble, white shirt with rolled sleeves and a slim navy tie) has lowered the can and holds the HOVER can at his chest; his eyes are wide open, eyebrows up, a stunned, euphoric smile spreading, as if a rush of happiness just hit his head; his hair and tie are lifted back by a soft wave of air and a few papers float up from the table. He stays seated. Warm real sunlight floods the boardroom. 28 mm. Same realistic cinematic photography as the cold scenes: ARRI Alexa, anamorphic lens, natural skin texture, fine film grain. Only the light has changed: warm late-afternoon sunlight now streams through the windows as a golden backlight, warmer white balance, richer natural colours in the room, crisp contrast. Realistic, not stylised: no fantasy sky, no glow or light effects, no oversaturation.
- Chain — built from the previous frame of the same take, so the room is identical.
- Performance — the body’s reaction, not a pose: no arms-up celebration.
- Product — the can stays in his hand, label to camera.
- Light — the same photography as the cold scenes; only the sun changes.
Images 1-8 are the storyboard frames, in order. Image 9 is the HOVER can: keep its design exactly. Video 1 is a grey 3D blockout: take only its camera movement and timing. Shot 1 (0-2 s), images 1 then 2: low camera at table height. The HOVER can slides into the frame from right beside the lens, crosses the table and stops upright next to Leo's limp hand. Leo stays asleep through the whole shot; his hand lies limp on the table from the first frame and does not move. Shot 2 (2-3.5 s), image 3: macro. His fingers pull the tab, a puff of cold mist. Shot 4 (5.5-9 s), images 5 then 6: ONE continuous take from a new low front angle, no cut inside it. He is still drinking, lowers the can and keeps holding it, swallows, opens his eyes, and a rush of happiness hits him … and the cold light warms into real sunlight. He stays seated, no jump. The can stays in his hand the whole time. Shot 5 (9-12 s), images 7 then 8: from behind his laptop lid. He leans back for a beat to collect himself, then snaps forward into the laptop, eyes darting, typing fast, charged with energy. Realistic warm light after the rush, no glow effects, no extra objects, never show a still photo. No subtitles.
- Frames in pairs — a start and an end state for every shot that changes.
- Motion reference — the Blender previs drives timing and camera.
- One take — the rush happens in the same shot as the sip, not after a cut.
- Sound — off; music, ambience and effects are built in the mix.
Finished in After Effects, not in the generator.
The master comp, exported from the real project: 31 layers, each named for what it is — every shot with its in and out, a grade per scene, the push into the can with its keyframes, the time remap that slows the last shot under the end card, and the sound stems. The film is rendered from this project, not stitched outside.
The cut
Trimmed, not regenerated
The can’s pop-in at the start of a take, a one-second freeze before the lob, a duplicate sip — cut out in the edit instead of spending another generation.
24 → 30 fps
Smooth, but never smeared
Pixel Motion for the calm shots. The fast ones — the push, the rush, the lob — go to 30 fps with repeated frames, so a flying can never melts.
Grade
Four scenes, one film
A cold, steel boardroom, warmer after the rush, the re-shot takes lifted to match their neighbours, and the final take warmed to sit with the morning office.
0:20
Into the can
A push into Leo’s laptop, a warm flash and a zoom blur that lands among the ice and beans.
End card
Logo on the final chord
HOVER and the line animate in their own precomp over a blur that builds while Tom drinks, timed to the last chord of the score.
Sound
Stems inside the project
Voice, music, ambience and effects sit in the comp as separate stems next to the master mix, ready to rebalance.
From 720p takes to a 4K master.
The video model delivers 720p. Every take is upscaled locally with Real-ESRGAN, blended 60/40 with the original so skin keeps its texture instead of turning to wax, then graded and given a fine grain in After Effects. Drag the line to compare.
720p take4K master
1:1 crop of the 4K frame · the take enlarged with plain bicubic
TakeFinal
Full frame · grade, upscale and grain
The music wakes up when Leo does.
The takes come out silent apart from one line, so the sound is built from scratch. Maya’s line is recast with ElevenLabs Voice Changer, which keeps the original timing, so her lips stay in sync.
The score is two cues: a sleepy muted-trumpet groove under the meeting that breathes through a filter on her sigh and sinks to a pulse when the can pops — and a bright one that drops on the exact frame the light turns warm. Five bars come out at the cut into the next morning, so its final chord lands on the logo. Every event has its own sound: 33 effects, each picked from four takes.
- 1Maya sighs — the groove breathes through a filter
- 2The can pops — the cue sinks to a pulse
- 3The light turns warm — the bright cue drops
- 4Five bars cut out at the jump to the next morning
- 5The final chord lands on the logo
Voice recast
Music cues
Sound effects
LUFS master
Blender
Previs: blocking, lenses, camera moves and timing for every shot.
GPT Image 2
Reference sheets, the can and the 23 keyframes.
Seedance 2.0
Video: one take per scene, guided by the previs.
ElevenLabs
Voice Changer, both music cues, ambience and sound effects.
Real-ESRGAN
Local upscale of every take from 720p to 4K.
After Effects
Edit, 30 fps conform, grade, transitions and the end card.
Contact
Want a film like this for your product?