I run a one-woman candle shop. Here’s what AI product videos actually cost me
I pour candles into stoneware vessels that a potter friend throws for me, and I sell them online. I can take a good photo in ten minutes on the kitchen table. Video is different. Every time I tried to film a candle being lit, I lost an afternoon to bad light, shaky hands and a match that wouldn’t catch.
This autumn I gave myself one week and a small budget to see whether AI video could fill that gap. Not for my product listings, and not instead of real photography. Just the short clips social feeds reward and I never have time to make.
My setup, and one honest caveat
I didn’t want to put my actual shop on display here, so I made three stand-in “product photos” with an image model: a speckled cream vessel on linen, the same vessel lit at dusk, and a shelf with three glazes. Each cost less than a cent.
The three stills I animated. All generated, no label or text on the vessels.
Then I tested two video models side by side, both called through the same account: Seedance 2.5 and Wan 3.0. I ran six clips on 19 November 2026, wrote down how long each one took and what it charged, and judged each against one question: could I post this today without editing it?
Test 1 and 2: Lighting the candle, my product photo in motion
This is the clip I’ve always wanted. I gave both models my cream vessel photo as the opening frame and asked for a hand to come in with a match, light the wick and pull away while the camera pushes in slowly.
Through the Seedance 2.5 API the five-second clip at 480p took 170 seconds and cost $0.48. The hand looks like a hand, the match flares the way real matches do, and the wax stayed the cream colour of my real candles. I’d post it as it is.
Wan 3.0 did the same job at the same resolution in 73 seconds for $0.16, a third of the price. The problem was the wax: as soon as the flame caught, the whole surface turned a warm peach. Nobody who buys from me receives a peach candle, so for a product clip that’s a no.
Same photo, same prompt. Top: wax stays true. Bottom: wax drifts to peach.
Test 3 and 4: A lifestyle clip from words alone
My best posts aren’t product shots at all. They’re moods. So I wrote one description of a woman in a knit sweater in a reading chair on a rainy afternoon, reaching for her tea, with a stoneware candle burning on the side table, and asked both models for an 8-second vertical clip.
Seedance 2.5 at 480p took 138 seconds and cost $0.77. The room and the rain were lovely, but my “stoneware candle” became a glowing amber jar that looks more like a salt lamp, and at one point her hand passed right over the flame on the way to the mug. I wouldn’t post a clip that shows customers doing that with a lit candle.
Wan 3.0 at 720P took 214 seconds, the slowest of my Wan clips, and cost $0.53. It gave me a matte stoneware candle that looks like something from my range, and she picked the mug up cleanly. I’d post it with an “AI-generated scene” note, because the woman isn’t a real customer and I won’t let anyone think she is.
Text-only lifestyle clips. The bottom one kept the candle looking like stoneware.
Test 5: Day to evening, from two photos
For this one I used the Wan 3.0 API with two images, my daylight photo as the first frame and the lit dusk version as the last, and asked for a fixed camera while the room darkens and the wick catches. Five seconds at 720P took 123 seconds and cost $0.33.
The light fades, the window turns blue, the flame appears and settles, and the vessel never moves. The wax warms toward amber in the last second, but because it happens with the evening light, it reads as lamplight rather than a different product. It’s now my favourite format, because my own photos control both ends.
First and last frame mode. Calm, accurate and cheap.
Test 6: The shelf, as a “final” clip
Last, I paid for 720p. I gave Seedance 2.5 my three-candle shelf photo as a reference and asked for a slow slide along the shelf, vertical, five seconds. It took 273 seconds and cost $1.09, the most expensive test by far.
All three glazes, the ridges on the green vessel and the terracotta lip survived, and the leaves at the edge moved gently. It did crop in tighter than my photo, so you never see all three candles at once.
Product details held. The framing got tighter than I asked for.
The numbers
Six clips, four I would post. Every task finished on the first try.
Six clips cost me $3.36 in total, plus about two cents for the stills. Four were good enough to post without editing, so that’s roughly 84 cents per usable clip. Every clip came back with a soundtrack by default; I muted them all and used my own music.
At the rates I was charged, Seedance 2.5 costs about 9.6 cents a second at 480p and 21.6 cents at 720p. Wan 3.0 costs about 3.3 cents at 480P and 6.6 cents at 720P. For lifestyle moods I now start with Wan. For anything where the product itself has to be exactly right, I pay more for Seedance.
My weekly routine now
On Mondays I pick two products and take one good photo of each, in daylight and at dusk. On Tuesday evenings I queue four clips while I pack orders, then check every frame, mostly for colour. Anything where the wax, the glaze or the size looks wrong goes in the bin, however pretty it is. Total time: about an hour a week, most of it spent checking.
What I still pay people for
AI clips are for my social feed only. My listing photos are still real photos of real candles. Anything showing burn time, scent throw or how a candle performs is filmed for real, because a generated clip can’t prove any of that. When I launch a new collection, I’ll still book a local photographer for a half day. And I don’t use AI faces as “customers” or testimonials, ever. Every AI clip gets a short note in the caption.
If you run a small product business and have been putting off video because of time, my advice is to start with your own photos as first and last frames. You keep control of what the product looks like, and you’ll learn quickly which clips are worth posting.



