DEV Community: Louis Chi

How to take product photos at home that AI video models can actually use

Louis Chi — Sun, 25 Oct 2026 11:58:43 +0000

Most "the AI ruined my product" complaints are photo problems. The logo melts, the bottle turns from teal to green, the shoe grows an extra stripe. Usually the input is one busy lifestyle shot, or five photos taken in five different lights.

I build RefVid, which turns product photos into video with reference-to-video models such as Kling Omni. We shoot many test sets for invented products (a serum called LUMEN, a shoe called AERO 1). This is what works with a phone and a kitchen table.

What a reference-to-video model does with your photos

Image-to-video animates one picture; once the camera moves past it, the model invents the rest, and products drift. Reference-to-video treats your photos as a description of the object and builds every frame of a new scene around them. The model can only keep what it can see, so it needs:

3 to 5 photos of the same object, not 1 and not 20.
Different angles of it: front, three-quarter, side, back, a detail.
The same light, background and color in every photo. The model averages across the set. If the photos disagree, you get the average.

That last point causes most of the failures, so the setup below is mostly about consistency, not beauty.

The setup: window, sweep, phone

Light. A big window without direct sun is the best free light you have. Put the product about a meter from it and stand a sheet of white card on the other side to lift the shadows. If you shoot at night, use two identical lamps at 45 degrees on either side. Never mix a warm bulb with daylight: the color shifts between angles, and the model learns a color your product doesn't have.

Backdrop. Curve a sheet of white or light gray paper (or bent foam board) from the table up to the wall so there is no horizon line. Set the product well forward on the flat part so its shadow doesn't climb the wall. Use mid-gray for white products so their edges don't vanish. No wood grain, no patterned cloth, nothing with text. A background that shows up in every reference gets treated as part of the product.

Light box. For small things (jewelry, cosmetics, earbuds), a light box gives soft, almost shadowless light from every side. You can make one: cut windows into three sides of a cardboard box, tape baking paper over them, line the inside with a curved sheet of white paper, and put one lamp on each side.

Phone settings. These matter more than the phone model:

Use the main lens at 1x or 2x. The ultra-wide bends straight edges, and the model will faithfully reproduce a bent bottle.
Turn off portrait mode, beauty filters and any color "look".
On iPhone, tap and hold the product until AE/AF Lock appears, then drag the exposure slider down a little so white labels keep detail. The stock camera can't lock white balance, which is one more reason to keep a single light type.
On Android, Pro mode lets you set white balance to a fixed value (around 5000 to 5500 K for daylight) and keep ISO low. Fix it once and leave it for the whole set.
Rest the phone on books or a tripod, use the 3-second timer, and wipe the lens.

The five angles (plus one)

Shoot these in order, without moving the lights:

Front, straight on at the product's eye level.
45-degree three-quarter view. Usually your best hero image, and the one we upload first, because the first image is treated as the main view.
Side.
Back, including the label panel if there is one.
Detail close-up of the logo or main marking, filling most of the frame and perfectly sharp.

The plus one: an in-use shot if the product is held or worn. A hand around the bottle, the watch on a wrist. Product-only photos don't show scale, and without one you can get a perfume bottle the size of a wine bottle.

One of the three plain-backdrop photos of LUMEN, a test serum. Same light, same paper, three angles.

Then pick 3 to 5 of these. Front, three-quarter and logo close-up is the minimum set we'd trust; add the back and the in-use shot if you have them. The full shot list, including top-down views and turntable clips, is in our guide to taking product photos at home.

Jewelry and other reflective things

Shiny surfaces mirror the room, and whatever is mirrored ends up in the video, in every scene. So:

Turn off the ceiling light and use one large soft source. Stand white card around the piece so it reflects plain white, and shoot from slightly above so it reflects the card, not you.
Shoot jewelry on mid-gray paper, or black for silver and diamonds. Wipe off fingerprints and dust first; a macro lens makes them huge.
Stand rings upright with a small ball of putty hidden behind the band.
Get as close as the phone still focuses (newer iPhone Pros switch to macro on their own) instead of using digital zoom.
Clear glass disappears on white. Put a strip of black card just outside the frame on each side to draw dark edges.

Once you have a clean set, a slow turn under moving light is where jewelry looks best; our jewelry sparkle template is that shot on black velvet.

Common failures, and the photo that fixes them

The logo warps or the text changes. Small printed text is the weakest point of every video model we've tested. You can't make it perfect, but you can stop making it worse: include one close-up where the logo fills the frame and is sharp at 100 percent zoom, and plan scenes where the logo is large and faces the camera. Fine print like ingredient lists will be redrawn whatever you do, so keep it out of focus or turned away. More on this in how to keep logos and packaging accurate.

The color drifts between clips. Almost always mixed light or mixed sessions: two photos from Tuesday's window, two from Friday's lamp. Reshoot all of them in one session, in one light. Also, shoot each colorway as its own set. A blue bottle and a red bottle in one reference set give you a purple-ish one.

The shape is wrong on the back or the sides. The model never saw them. Add the side and back photos. A product with a distinctive cap needs a top-down shot too.

Random props show up in the scene. They were in the references. Crop them out, and keep lifestyle photos out of the set.

The prompt fights the photos. Once the photos are attached, don't describe the product again in words ("a teal glass bottle with a gold cap"). Name it and describe the scene instead. When words and photos disagree, the model blends them.

From photos to clips

AERO 1, an invented running shoe: three photos against one gray backdrop, then a gym scene generated from them with Kling 3.0 Omni.

Once the set is right, the video part is short: save the product once, write one line about the scene, and every clip starts from the same object. Details, including what to do with only one shop photo, are in product photo to video.

The clips worth making first are the ones a shop actually uses: a 360-degree turntable for the product page, a vertical clip of someone holding it for TikTok and Reels, an unboxing. Our ecommerce video guide covers which format goes where.

The checklist

One light type, no mixing.
Plain curved backdrop, product well forward.
Main lens, no filters, exposure locked.
Front, three-quarter, side, back, logo close-up, plus one in-use shot.
Colors match across the set; one set per colorway.
Logo sharp at full size, checked on a big screen, not the phone.
Hero angle uploaded first, 3 to 5 photos in total.

Side effect: these are the photos every shop wants anyway, so one afternoon at the window covers both the listing and the video. If your output still drifts, tell me in the comments what broke.

How we keep one character's face the same across AI video shots (Kling reference-to-video, with code)

Louis Chi — Sat, 24 Oct 2026 13:59:29 +0000

If you've built anything on top of a text-to-video model, you've hit this: the first clip is great, the second has a slightly different nose, and by the fifth it's a different person. A prompt like "a woman with a copper-red bob and an olive jacket" fits millions of people, and the model picks a new one on every request.

I run RefVid, a small tool built around one idea: you save a character or a product once, as a few photos, and every clip after that is generated from exactly those photos. This post covers what that takes in code. We use Kling's reference-to-video endpoints through fal's queue API, and everything below comes from our production code and our own test clips.

Words alone drift; photos hold

We started by measuring the problem. We took one invented character, Mia, and generated the same three scenes twice on Kling 3.0 Omni, 5 seconds each:

Prompt only, with Mia described in as much detail as we could fit (hair, freckles, earrings, jacket).
The same scene prompts with no appearance words at all, plus four reference photos of Mia attached.

Prompt only. A copper-red bob and an olive jacket, but not quite the same woman as in the other scenes.

Four reference photos, no appearance words in the prompt. The same face in every scene.

The prompt-only set cost $1.26 for three clips, and each one gives you a different person who matches the description. The reference set keeps her. The full side-by-side is on our consistent character guide.

What we took from it: once you attach references, stop describing the person in the prompt. When the words and the photos disagree, the model blends them, and the blend is exactly the drift you were trying to fix.

The request shape: elements, not just images

Kling's reference-to-video takes subjects as elements. An element is one frontal image plus extra views of the same subject. The prompt points at it as @Element1, @Element2, and so on. Here is the core of how we turn a saved subject into an element (simplified from our provider):

const MAX_ELEMENT_IMAGES = 4; // one frontal + up to three extra views per element

function toElement(imageUrls: string[]) {
  const [frontal, ...rest] = imageUrls;
  if (!frontal) throw new Error('at least one reference image is required');
  const views = rest.slice(0, MAX_ELEMENT_IMAGES - 1);
  return {
    // reference_image_urls needs at least one image, so a one-photo subject repeats its frontal view
    element: { frontal_image_url: frontal, reference_image_urls: views.length ? views : [frontal] },
    // photos beyond four go in as general references instead of being dropped
    extra: rest.slice(MAX_ELEMENT_IMAGES - 1),
  };
}

Three things we learned the hard way:

Order matters. The first image is treated as the frontal view, so make it your sharpest, most evenly lit, straight-on shot. A three-quarter view in that slot gives you a slightly turned face in every clip.
One photo still works, but it has to be sent twice. The endpoint wants at least one extra view, so a single-image subject sends its frontal again.
Overflow goes to image_urls. If someone uploads five photos, the fifth still helps as a plain reference instead of being silently ignored.

The prompt must name every element

If the prompt never mentions @Element1, the model is free to ignore the photos. People don't write @Element1, though. They write "Mia pours coffee". So before a prompt goes out, we swap each subject's name for its token:

// "Mia holds NOIR 07 and says \"Hi, I'm Mia\""
//   → "@Element1 holds @Element2 and says \"Hi, I'm Mia\""
bindSubjects(prompt, ['Mia', 'NOIR 07']);

The details are where it breaks:

Whole words only. "Mia" must not match inside "Miami". We use Unicode-aware boundaries, (?<![\p{L}\p{N}]) before and (?![\p{L}\p{N}]) after. They're skipped for Chinese, Japanese and Thai, where words have no spaces around them.
Not inside quoted dialogue. Kling 4 clips have sound, and if the swap reaches the spoken line, the voice literally says "at Element one". Text inside "…", “…”, „…“, «…» and 「…」 is left alone.
Unnamed subjects get prefixed. If a subject is attached but never named, its token is put at the start of the prompt so the model still uses it.
Respect authors who already use tokens. If a prompt already contains @Element1, we leave it untouched.

Long sequences: split, then chain

Faces hold up best in short clips. Drift creeps in as the camera moves and the face turns to angles the references don't cover. So instead of one long generation, we do two things.

Multi-shot requests. Kling accepts a multi_prompt list, one prompt per shot, and splits the duration between them. On Kling 4, each shot's prompt is capped at 512 characters. Names and translations can push a shot over that cap, so in that case we send one prompt that lists the shots in order rather than let the request fail:

const shotsFit = shots.every((s) => s.prompt.length <= 512);
const body = shotsFit
  ? { multi_prompt: shots, elements }
  : { prompt: shots.map((s, i) => `Shot ${i + 1} (${s.duration} s): ${s.prompt}`).join('\n'), elements };

Chained clips. For a series, we start each clip from the last frame of the previous one (start_image_url) and attach the same element every time. We made a four-shot running-shoe sequence this way: four 5-second clips at $0.42 each, plus $0.03 to pull the last frames out.

Shot 1. Each of the next shots starts from the previous shot's last frame, with the same product element attached.

Shot 4. Same shoe, same path, same light direction.

The start frame carries the scene and the light. The element carries the identity. You need both: with only the start frame, the identity slips once the camera moves away from what that frame showed.

Products are harder than faces

People forgive a slightly different jaw. Nobody forgives a misspelled logo. For products:

Shoot three photos on a plain backdrop: the hero angle, a three-quarter view, and a close-up where the label fills the frame.
Put the label in the prompt as a fact about the scene, not as a description: "the LUMEN label faces the camera and stays readable".
Keep the label large in frame in at least one shot. Small text in a wide shot has too few pixels to come out right.

LUMEN is a serum we invented for testing. Three reference photos, Kling 3.0 Omni; the label carries over into each scene.

Plumbing notes

Use the queue and a webhook, not a blocking call. A 3-second 720p Kling 4 clip with sound took about 80 seconds to generate for us. We submit to queue.fal.run/<model>?fal_webhook=<url>, verify the webhook signature, and keep a status poll as a fallback.
Treat a 5xx on the result fetch as retryable. A failed fetch of the result doesn't mean the job failed. We only mark a generation failed when fal reports an error for the job itself.
Make the webhook idempotent. Record the request id before you act on it. Webhooks get delivered twice, and you don't want to refund or charge twice.

What I'd tell anyone building this

Version the reference set. Most "the face changed" bug reports come down to clip 1 and clip 4 having been made from different photos.
Strip appearance words from prompts once references are attached.
Keep clips short and cut between them. Chaining 5-second clips beats one 15-second clip for identity.
Test with an invented character, so you know what "the same person" means.

If you want to try it without building any of this, RefVid does all of the above behind one upload. Ask in the comments about the request shapes or the regexes. I'm happy to go into more detail.