Indoor skydiving: first-last empty tunnel to airflow, not a generated flyer

Two stills of the same chamber. Flux 3 interpolates fan. generate_sound_effect for the roar. Do not seat a generated body in the column.

A tunnel that books is the flight chamber you operate: glass, mesh floor, the padded ring, the empty column. It is not a generated flyer in a suit. Those faces are private people the model invented, often minors. Photograph the empty chamber twice — fans off, then airflow you can see. Interpolate the fan. Lay the roar as sound you generated. Leave the body out of the column.

Two stills of the same empty chamber

First-last frame is a two-still contract. Ordinary image-to-video knows where it begins and improvises the rest. First-last has to land. Empty chamber, still air is the start. Empty chamber, visible airflow is the destination. The model is only allowed to travel between those two photographs.

Photograph the tunnel you sell. Same lens height a visitor would have through the glass. Same padding, same mesh, same logo in the ring if it is really there. Frame A: fans down, no haze, no people. Frame B: the same marks, the same glass, airflow you actually produce — haze, a streamer — still no people. Do not composite a “busier” tunnel from a different facility.

The two frames have to be connectable. Same room, same rotation, same distance to the lens. A wide empty chamber morphing into a macro of a fan grate is a warp, not a spin-up. Keep the whole column. If the endpoints disagree about padding or a logo, the middle will invent a third tunnel.

Crop people out of both stills before you generate. A reflection, a staff member, a child at the window — strip those in the photograph, not in a prompt. Do not ask a model to “remove the flyer and keep the column.” It will rebuild both.

Create a first-and-last-frame transition video is the named agent job: two photos in, one transition out. Cost is priced per model, shown before you confirm. This is not the plain “turn one photo into a video” capability. Missing the airflow still is a different job.

Flux 3 interpolates the fan

Flux 3 First Last Frame to Video is the catalog row for that interpolation: start frame and end frame, native audio, 720p or 1080p, durations from 5 seconds through 20. The listed rate is 9 credits per second. Audio is on — write the mix or write silence. A silent brief still gets a soundtrack you did not choose.

Prompt the travel, not the occupants. The stills already contain the glass and the mesh. Name the mechanism: hold, then the column fills with visible air; camera locked; no one entering frame. If the endpoints share geometry, the morph reads as a spin-up. If they do not, you get a mushy crossfade.

Do not re-describe “a flyer hovering in the tunnel.” Flux 3 will try to grow one in the middle. If a body appears, throw the take away. Five seconds is a short spin-up. Ten is a readable column. Twenty is the ceiling. Until the chamber holds, stay on the short end.

Native audio rides along. That is not the roar you want to control. Write silence in the brief — no speech, no music, no hoped-for fan — or you will get a soundtrack you cannot edit as a stem. The interpolate is the picture. The roar is a different job.

The roar is generate_sound_effect

Generate a sound effect is the named agent job: generate_sound_effect from a prompt, optionally with duration and loop. Describe the tunnel roar. Get the SFX. Place it. Slip it a frame if the spin-up is late.

This is not a song. A whoosh is generate_sound_effect, not a song prompt: generate_music writes a track. Generate the roar. Place it. Regenerating the video so the picture matches a baked-in ambience is the expensive diagnosis.

Say whether it loops. Say how long. Cost is per generation, shown before you confirm. Do not ask Flux 3 to “add the real tunnel sound.” Sync is an edit: nudge the start.

A constructed instructor explaining “what to do if you drop a shoulder” is a health-adjacent persona. You do not need that mouth. The empty column plus the roar is the ad.

Do not seat a generated body in the column

A deepfake of a private person is not a style. TikTok’s AI-generated content rule is explicit about young people under 18 and adult private figures used without permission. Labelling the clip AI does not create a grant. Keep people out of the prompt. Keep people out of both plates.

Indoor skydiving ads fail when they recast the chamber as a packed class. First-timers are often minors. A generated pack of smiling kids in the column is the under-18 line. “Family fun in the tunnel” is how those faces appear.

Do not prompt a photoreal flyer in the last frame. Do not treat a loop (same still in both slots) as a spin-up. A loop is idle motion on one plate. This offer needs two states.

If a session needs a real named flyer, that is a filmed day with consent. Generation can still do the fan. It cannot be the body. The chamber is the SKU. The roar is the SFX. The flyer is not yours to invent.

FAQ

Can I animate one photo of the empty tunnel and hope the fans spin up?

You can. The ending is then the model’s guess. The product is the airflow you actually run. Pin that still as the last frame or you are not selling the column.

Why generate_sound_effect instead of Flux 3’s native audio for the roar?

Because native audio on Flux 3 is a mix baked into the take. A silent brief still gets a soundtrack you did not choose. Generate a sound effect is a stem you can place, loop, and slip. Write silence on the interpolate. Lay the roar after.

Can we generate generic happy flyers if they are not real customers?

No. Invented faces are still likenesses, and invented children are the under-18 line. Sell the empty chamber. Generate the roar. Do not seat a body in the column.

Do the two photos have to be the same tunnel?

Yes. Same chamber, same lens family, same lighting neighbourhood. A different facility is a morph. The agent job will not rescue a mismatched pair.