Slit-lamp ads are the empty chinrest, never a generated eye
Photograph the tower and empty chinrest; GPT Image 1.5 (4K, 2cr) holds the unit plate and overlay magnification as type — a photoreal patient or talking optometrist is a YPP kill, so do not I2V a face into the rest.
Photograph the tower and empty chinrest; GPT Image 1.5 (4K, 2cr) holds the unit plate and overlay magnification as type — a photoreal patient or talking optometrist is a YPP kill, so do not I2V a face into the rest.
A slit-lamp listing a buyer can match is this tower, this empty chinrest, the turret you ship. It is not an eye, not a constructed clinician walking “what you should expect from the exam.” Photograph the unit empty. Generation may rake the bench light. It may not invent a body.
Photograph the tower and empty chinrest
Procurement already knows the category. What they need is the serial they quoted: binocular head, slit projector, chinrest height. Text-to-video of “a modern eye exam, photoreal patient at the slit lamp” invents a cousin tower and a face you did not consent.
Shoot the approved unit like a line card. Square-on or a three-quarter that still reads the chinrest and the turret. Chinrest empty. Forehead strap parked. No extra, no dummy iris, no “anonymised” cornea. If the only photo includes a face, crop them or shoot again.
Text-to-image is not the site plate. A paragraph will invent a nicer bench. Prompt later motion only as camera on the file: a slow orbit of the tower, a hold on the empty rest. A generated eye is anatomy. Do not prompt a dilated pupil, a retinal close-up, a child’s eye, or a “model cornea for scale.” Empty rest, or no people.
GPT Image 1.5 holds the unit plate
GPT Image 1.5 is OpenAI’s row across text-to-image, image-to-image, and edits. It is strong at long, specific instructions and legible in-image text. Content type is image. Max output is 4K. Display price is 2 credits. Audio is null. No duration: this is not a clip.
Use it when the unit plate must be drawn or repaired: model string, the turret labels you actually print. Quote the string. Proof at 100%. A wrong digit is a different machine.
If the plate is already sharp in a photo you took, keep the photograph. GPT Image 1.5 is for type that is not in the file yet, or for an edit on a locked still. A fresh text-to-image is a new sample. It will be 4K and close, and it will not be the file you approved.
Aspects include 1:1, 2:3, and 3:2. Qualities include SD, HD, and 4K. Stay on 4K if the locked file is 4K. Magnification numbers that have to survive a pause belong on this still, or they belong off the generate as overlay.
Overlay magnification as type
10×, 16×, 25×, 40× — those figures are a turret you already ship. Do not ask a video model to letter them into the drum. Glyphs a generate draws have to be right every frame. They will not be.
Add a text overlay burns one authored line for the whole clip — top, center, or bottom. That is the magnification you sell, the model string, “for demonstration only.” It does not transcribe. Write the string from the IFU. Picture lock, then type. Crop to the ship ratio before you overlay.
Auto-captions listen. An empty chinrest has nothing to transcribe. A silent tower on a real bench is typesetting, not a mouth. Keep lips out of frame. If someone must speak a protocol, that someone is a real named credentialed person on a different take. Generation can still do the tower. It cannot be the consult, and it cannot be the eye.
Do not I2V a face into the rest
When frame one has to be the JPEG you approved, open image-to-video on that file. The family you run through it is Seedance 2.5: 4s–30s in one pass, 720p ceiling, listed 47 credits, native audio. I2V is the still-to-motion job on that family. The published URL is this slug. Do not invent a second Seedance path.
Iterate at 4s or 5s until the empty chinrest holds. Native audio is on. Write a quiet suite or silence. A silent brief still gets a soundtrack you did not choose.
Do not I2V a face into the rest. A still that already has a patient will grow a patient. “Add a photoreal sitter so the chinrest makes sense” is anatomy you did not consent.
Synthetic presenters on health and finance are a YPP kill. Help’s third inauthentic-content bucket is AI personas that present as human experts on health. A constructed optometrist walking “here is what you should do before dilation” is that shape. Labelling the face AI does not leave the bucket.
Pause on the empty rest. If an eye, a face, or a clinician appeared who was not in the still, you made an exam film. Crop them or reshoot.
FAQ
Can we generate an eye so buyers see the beam?
No. That demo is anatomy. Photograph the parked tower and empty chinrest. Overlay magnification. Do not invent a cornea so the unit has a scene.
Why not a talking optometrist walking the exam?
Because a photoreal talking presenter on health is YouTube’s third inauthentic-content bucket. Disclosure does not move a constructed clinician out of it. If someone must speak, put a real named clinician on camera.
Do I still need GPT Image 1.5 if the plate is already in the photo?
No. Keep the photo. GPT Image 1.5 is for type that is not in the file yet. Animate the real tower from that file.
Can I I2V the empty chinrest?
Yes, from the empty still, via image-to-video on the Seedance 2.5 family: 4s–30s, 720p, 47 credits, native audio. Prompt camera only. Do not I2V a face into the rest.