SpO2 probe clips render in Imagen 4 (4K, 4cr, text-only); never a generated fingertip

SpO2 clip-probe SKUs render in Imagen 4 at 4K, 4cr, text-only, then overlay the wavelength; never a generated fingertip, never a photoreal talking clinician, and the model cannot ingest a patient still.

An SpO2 clip listing a buyer can match is this probe: jaws, spring, emitter window, cable strain relief. It is not a generated finger in the clip, and it is not a photoreal clinician walking a saturation lecture. Prompt the empty hardware. Burn the wavelength string you already publish. Leave tissue and the mouth off the generate.

Prompt the empty clip, never a fingertip

A clip-probe is a SKU with a serial, a wavelength pair, and a jaw a clinic will match to a quote. Text-to-video of “a pulse-ox clip on a patient finger in a busy bay” invents a prettier cousin and anatomy you did not consent. That is a different probe and a patient you did not hire.

Photograph the unit you sell if you already have it — clip parked, jaws empty. If the only photo includes a finger, crop until only plastic and cable remain, or shoot again. Do not prompt the anatomy out and hope the model rebuilds the clip. Do not prompt a fingertip “for scale.” A fingertip is a person. Scale is a known jaw width, a tape, the clip on a table.

Patient monitors: bezel still; waveforms are overlay, not Veo vitals already spent the chassis: authored traces, no talking white coat. Do not copy that page onto a clip. The leftover here is the probe, not another demo screen.

Imagen 4 is 4K, 4 credits, text-only

Imagen 4 is Google’s text-to-image row: 4K, 4 credits, photoreal, text-only. Category is text-to-image. audio is null. No duration. It cannot ingest the probe photo. It cannot ingest a patient still. A paragraph is the whole brief. A finger the prompt implied is not yours.

Use Imagen 4 only when you need a closed, empty SKU still of the clip you sell — jaws parked, window empty, no body, no clinician — and you will label the picture as a product render, not a ward. Proof at 100%. A wrong jaw is a different SKU. Do not send Imagen a paragraph that includes an exam. “Photoreal SpO2 clip on a fingertip, 98%” decides anatomy and a reading you did not take.

Generate an image from text is the named agent job for that paragraph: a prompt, an optional ratio, a still. Cost is priced per model, shown before you confirm. Name Imagen 4 when the brief is this 4K, 4-credit, text-only plate. The text-to-image tool is the door.

When frame one has to be the JPEG you approved, skip Imagen. This row will not take the file. It will decide a clip that is not the serial on the work order. Do not ask Imagen to letter wavelength into the housing, or to paint a 98 on a window. That digit is a vital. Vitals are claims.

Overlay the wavelength; check the contrast

Wavelength is a claim from the cut sheet you already ship. Red and infrared pairs — whatever this SKU actually fires — are why a buyer opens the listing. A 660 nm that becomes a cousin band because the model liked a red glow is a mis-ship. Video models do not typeset it.

Add a text overlay to a video burns one authored line for the whole clip — top, center, or bottom. It does not transcribe. Write the string as it sits on the IFU: wavelength pair, a SKU, “for demonstration only.” Overlay after picture lock. Crop to the delivery ratio before you overlay. If the still already carries a sharp legend you photographed on the probe you ship, keep it. Do not “enhance” the badge.

Contrast Checker is the arithmetic for that print: two hex values in, a WCAG ratio out. Foreground is the overlay colour. Background is the clip plastic or the scrim you will actually paint. The ratio is (L1 + 0.05) / (L2 + 0.05) with WCAG 2 relative luminance. Below 18pt regular / 14pt bold, 4.5:1 (AA) applies. The checker does not upload a screenshot and does not spend a credit. Measure the pair you will burn.

Never a photoreal talking clinician

Synthetic presenters on health and finance are a YPP kill. Help’s third inauthentic-content bucket is AI-generated personas that present as human experts on health. A constructed respiratory therapist or clinician telling a viewer what a saturation “should” read is that shape. Labelling the face AI does not leave the bucket.

Do not glue a talking white coat onto the clip. If a protocol needs a voice, that voice is a real named credentialed person on a different day. Generation can still do the probe. It cannot be the clinician.

A generated fingertip is anatomy. Do not prompt a finger in the jaws, even “anonymised,” a teaching-file still “for scale,” or a paediatric extra. A child-sized finger is a hard stop. A silent empty clip on a table is not the YPP shape. If the first three seconds look like an exam, rewrite until they look like a clip.

FAQ

Can we generate a fingertip in the clip so buyers see how it seats?

No. That field is anatomy. Photograph or render the empty clip. Overlay wavelength and the IFU URL you already host. A real named applications person, on a different shoot, if you need a voice. Do not invent tissue so the probe has a scene.

Can I upload a clinic photo into Imagen 4 so the clip matches?

No. Imagen 4 is text-only, 4 credits, 4K. It does not take a reference still. It cannot ingest a patient photo. Prompt the empty SKU, or photograph the unit you actually own and keep that file off this row.

Is a disclosed AI clinician acceptable on the product channel?

No. Disclosure does not move a constructed health expert out of YouTube’s third bucket. Keep the empty clip. Keep the overlay. If someone must speak a protocol, that someone is a real named credentialed person, not a generate.

Why use Contrast Checker instead of asking Imagen to pick a readable colour?

Because contrast is a formula, not a prompt. Contrast Checker takes two hex values and reports the WCAG ratio. Imagen will invent type and a colour that looks fine in a bright room and still fails 4.5:1. Measure the pair you will burn. Do not generate the arithmetic.