This afternoon, recording a tutorial for a feature we'd just shipped, I pressed Remove old photos, watched two flat T-shirt photos leave the row, opened a video template, and there they were again.
The button had done exactly what the code told it to. It changed what was on screen and never touched the saved product.
That button is the last step of a feature we built today: turning flat product photos into four studio views before a video is made. Here's why the feature exists, how it works, and the two state bugs it shipped with.
TL;DR
A video model given one flat photo of a shirt has to invent the back and sides when the clip turns it. So RefVid now draws four studio views (front, back, left, right) from your flat-lay photos first. On day one, removed photos came back on reload, and a closing dialog briefly announced "Your 0 new views are ready."
A flat photo shows one side, and a turning video needs four
A lot of clothing gets photographed the cheap way: laid flat, shot from above. That's a fine photo for a product page.
It's a poor photo for a video. Ask a model for a slow turn or a floating spin, and halfway through it has to show a side it has never seen. It doesn't know what the back print looks like, how the sleeve falls, or whether there's a print at all. So it makes something up, and the made-up back is what the viewer sees for half the clip.
A prompt can say "white tee, logo on the back," but it can't say what the logo looks like. The model needs pictures of every side. Drawn views can't recover details your photos don't show; what they do is give the video model one consistent shirt from four angles instead of one flat one.
So the studio draws the views first
I build RefVid, a small tool where you upload a few photos of a product or a character once, then make short videos that keep it looking the same in every shot. Today we shipped a step that sits in front of the video.
When a product's main photo is in the composer, a small wand sits on its corner:
Clicking it opens a card. You tick up to three source photos (the first ticked one is treated as the front) and the views you want. The button spells out the total before you press it:
The whole flow, start to finish:
| Step | You do | What happens |
|---|---|---|
| 1 | Add the flat photos to a product, front first | They become the product's photos |
| 2 | Click the wand on the main photo | The Draw product views card opens |
| 3 | Tick source photos and views | The button shows the total, e.g. Draw 4 views · 120 credits (30 a view) |
| 4 | Press Draw | About a minute later the views take the product's first places |
| 5 | Press Remove old photos | The flat photos are deleted from the product |
| 6 | Pick a template (floating product, 360 turntable) | The video starts from the four views |
The views are ghost-mannequin shots: the shirt filled out as if someone were wearing it, nobody inside, floating over light grey. Underneath is an image model; RefVid writes its instructions. A view is charged once it's stored, and a view that fails gives its credits back.
The sides are drawn from the new front, not the flat photo
Four views drawn independently would be four independent guesses. The front might get a slightly boxier cut, the left side a slightly different white, and a turning video would stitch those into a shirt that changes shape mid-spin.
So the drawing runs in two rounds. First the front and back, straight from your flat photos. Then the left and right, with the front that was just drawn as their first picture. The sides never see a wrinkled shirt on a floor; they copy a studio shot that already has the light, the grey and the cut decided. The back doesn't need the front (your back photo already says what's there), so two rounds are enough, not three.
Here is the result on our test shirt: flat photos on the left, the four views on the right, and a close-up of the back print.
The delete button changed the screen and left the saved photos alone
Once the views exist, the flat photos get in the way. If they stay in the row, the video model gets both a crisp studio front and a wrinkled shirt on a wooden floor: two versions of one shirt to reconcile. So when the views land, RefVid asks: Happy with them? Remove the old photos from this product…
The first version of that button removed the flat photos from the clip being composed. The row looked right.
But the composer doesn't own a product's photos. Open a template link or reload, and it rebuilds the row from the photos stored on the product, which nobody had touched. The flats came straight back, and the next video would have started from all six pictures: the exact mix the button was meant to prevent.
The same-day fix makes the button mean what it says: it deletes the old photos from the product itself, the same as pressing × on the product's page. One guard went in with it: nothing is deleted unless every new view is already stored, so a half-finished drawing can't leave a product with no photos.
When a button's label is a verb, check which copy of the data the verb reaches. "Remove" in a composer and "remove" on a product look identical on screen.
The dialog counted views that were already gone
The second bug: for a fraction of a second, as that dialog faded out, its title read "Your 0 new views are ready."
The count came from the same state that told the dialog to be open. Closing cleared that state at once, but the fade-out was still running, so the dialog rendered its last frames with nothing in them. It now keeps the views it was showing until it has fully gone.
What it can't do
- Very small text and busy all-over patterns come out softer. A big logo holds up; a care label or a dense print doesn't always. Check each view against your photos and redraw the one that's off (30 credits for one view).
- Products that aren't worn float. A bottle, a shoe or a box is drawn upright and floating in the same grey studio. Fine for a turntable; not a lifestyle shot.
- No back photo, no back print. Upload only the front and the back is drawn plain in the shirt's colour. It doesn't invent a print you didn't show it.
- The example is synthetic. The tee in every screenshot is a RefVid logo tee, and its "flat-lay photos" were themselves made with an image model, so this walkthrough doesn't show how it does on a photographed garment.
- Look at the left view above. It has thin dark marks along its edges that the source photos don't support; a shirt seen exactly side-on wouldn't show them.
Try it
The walkthrough, with a screenshot for every step, is on the tutorial page, and there's a 2:43 narrated video if you'd rather watch.
Which button in your app changes what's on screen but leaves the saved data untouched, and how did you find out?
Written with AI assistance from our own build notes and tickets.




Top comments (0)