Product Videos for a Store Catalogue With Seedance 2.0 Fast: Nine Tests on Colour, Labels and the Back of the Bag
Frames from the nine generated clips. All of them started from still images of products that exist only for this test.
A short video on a product page shows things a photograph cannot: how a bag holds its shape from the side, how light moves on waxed canvas, what is on the back. For a small store, shooting that for every item and every colourway is the problem. A photographer, a turntable and an afternoon cover perhaps a few dozen products. A catalogue may have several hundred.
Video generation models look like the obvious shortcut, but a product video has a stricter job than most generated video. A travel clip only has to look good. A product clip has to show the item the customer will receive, with the right colour, the right number of pockets and the right words on the label. If it invents a feature, you may end up handling a return or a complaint.
So we tested accuracy, not beauty. The store is fictional: Tarnwick Supply, an outdoor goods shop we invented for this article, selling a waxed canvas backpack in three colourways and a tin of coffee. We ran nine tests with ByteDance's Seedance 2.0 Fast (seedance-2.0-fast), using nine clips and nine still images for $1.85 in generation fees. We measured colour, read labels with OCR, and looked closely at sampled frames.
Two limits up front. The input images are generated stand-ins at $0.003 each, which is why the Tarnwick name renders so cleanly. Your own photographs will behave differently, particularly if they are less tidy. And each result is a single clip, with one seed, so a pattern here is a lead and not a statistic.
The setup
The API is asynchronous: you post a job, receive a task ID, poll until the job completes, and download an MP4. The code below is a simplified version of our runner, not the exact code we ran, and it uses a placeholder host. The real runner also logged each task ID to disk before polling and retried failed polls.
import os, time, requests
BASE = os.environ["VIDEO_API_BASE"] # placeholder host
H = {"Authorization": "Bearer " + os.environ["VIDEO_API_KEY"]}
job = {
"model": "seedance-2.0-fast",
"prompt": PROMPT,
"image_urls": [PRODUCT_PHOTO_URL],
"duration": 5,
"resolution": "480p",
"size": "adaptive",
"seed": 5150,
"generate_audio": False,
}
r = requests.post(f"{BASE}/v1/videos/generations", headers=H, json=job, timeout=60)
task_id = r.json()["data"][0]["task_id"]
save_to_database(task_id) # before polling, so a crash never pays twice
while True:
time.sleep(15)
t = requests.get(f"{BASE}/v1/tasks/{task_id}", headers=H, timeout=30).json()["data"]
if t["status"] in ("completed", "failed"):
break
Every clip was 5 seconds, seed 5150, with no audio. All but one were at the 480p tier. Square inputs came back as 640x640 at that tier and 960x960 at 720p, in H.264 at 24 fps. The default prompt for the backpack tests was:
Slow camera arc of about 20 degrees around the product. Soft light sweeps across it. The product keeps its exact shape, colour and details, including any printed or woven text. Background stays seamless light grey. No people, no added objects, no added text.
Test 1: describing the product or showing it
The cheapest possible catalogue workflow would feed product descriptions straight into the model, with no photography. So we ran one clip from text alone, using almost the same words that generated our reference image, and one clip from the image. Each cost $0.18. Render time was 99 seconds for the text clip and 132 for the photo clip.
The text clip has plausible parts and a correct patch, but it is a different product.
The result was instructive because the obvious measurement missed it. Colour was not the problem: the median product colour in the text clip stayed a mean Delta E of 4.5 from the reference, against 4.6 for the photo clip. OCR read TARNWICK correctly in all 30 frames we sampled from the text clip. A checker that looked only at colour and text would have passed it.
But the bag is wrong. It has no front pocket, the marbled wax texture has become smooth canvas, the straps carry rivets the reference does not have, and there is a contrasting base. All of these were things the prompt left open, and the model filled the gaps. For catalogue work, this settles the question: text-only generation is for mood and lifestyle shots, and anything that represents a specific item needs a photograph as input.
Test 2: where does the clip actually start?
If you want a video to open on your approved product shot, for instance so that it can replace a thumbnail without a visual jump, the first frame matters. We compared the first frame of each clip with its source image using SSIM, where 1.0 means identical.
Same input photograph. The prompt and the input method both change the opening frame.
Passing the photo in image_urls does not pin the opening. With the camera arc prompt, the green backpack clip opened already rotated, at an SSIM of just 0.64, and with the same prompt and seed the rust and navy clips scored 0.87 and 0.92. With prompts that held the camera still or spun the product, image_urls gave 0.91. The explicit option, image_with_roles with a first_frame entry, gave 0.95. The two tin clips, with a very slight push-in and a flat simple subject, reached 0.96.
The practical rule: if the opening frame matters, use the first_frame role, and do not ask for camera movement in a way that invites the model to start somewhere else.
Test 3: three colourways, one prompt
Stores with colour variants would want one prompt to work across them. We ran the same prompt and seed on green, rust and navy photos, at $0.18 each. We measured the median colour of the product in 11 evenly spaced frames, separating it from the grey background, and compared that with the same measure on the source photo using CIE76 Delta E. As a rough guide, a difference of 2 or 3 is around where a careful viewer starts to notice, and above 10 is plainly a different shade.
The same prompt held the navy bag almost exactly and took the rust bag noticeably darker and redder.
|
Colourway |
Mean Delta E |
Largest |
What we saw |
|
Navy |
1.6 |
2.4 |
Very close, but the lower pocket turns visibly paler around 2.5 s |
|
Green |
4.6 |
12.4 |
About 10 to 12 in the first half second, then under 5 from about 1 s on |
|
Rust |
17.0 |
24.5 |
Darker throughout, 15 lightness points down by the end |
Some of this is lighting and not colour error, because we asked for a light sweep and the background shading moves as well. But a customer does not care why the rust bag looks brick red. In this one run the saturated orange fared much worse than the dark navy, so bright or saturated variants are where we would look first. Note too that the navy bag scored best on the number and still had a visible fault, the pale pocket, which the median did not catch. Numbers narrow the review, but somebody still has to watch the clip.
A fix we would try first, though we have not tested it, is to drop "light sweeps across it" from the prompt for any variant where colour is the selling point.
Test 4: can the label be read?
A label is a good trap for generators, and a net weight is a fact a customer relies on. Our tin had three lines of text: TARNWICK, TRAIL ROAST and 340 g NET WT. The prompt was:
Very slow push-in toward the label. The tin stays upright and still. The label text stays sharp and unchanged. Background stays seamless light grey. No people, no added objects.
We ran it at both tiers: $0.18 at 480p and $0.39 at 720p. Then we extracted six frames per second, 30 per clip, doubled their size and ran them through the OCR in Apple's Vision framework.
The 720p letter edges are cleaner, but both are readable, and OCR agreed on every sampled frame.
All 30 frames read exactly "TARNWICK / TRAIL ROAST / 340 g NET WT." at both tiers. For a large, high-contrast label with a static subject, the cheaper tier was enough, and the extra spend bought cleaner edges only. On the backpack, whose patch is much smaller in the frame, the arc clip gave 27 exact reads out of 30, with two stray punctuation marks and one misread, TARNWCK.
Do not stretch this result to ingredient lists or care labels. Our label was bold and took up a large part of the frame, and we did not test fine print. We would treat any regulated or safety text as something to overlay on the finished video, or to keep on the page next to it.
Test 5: a full turn from one photo
This is the shot many store owners would ask for first: a 360-degree spin. We gave the model only the front photo.
The backpack rotates a full 360 degrees at constant speed on an invisible turntable and ends facing front again. Static camera. Background stays seamless light grey. No people, no added objects.
Convincing motion, invented product. Compare the shoulder straps on the left with those in the clip.
As motion, this was impressive for $0.18 and 148 seconds of rendering. The bag completed a full turn, returned to the front, and held its colour with a mean Delta E of 3.3. But the back is fiction. Our rear still shows padded tan leather shoulder straps. The clip gives the bag green canvas straps ending in tangled tan loops. No model can know what is behind the photo, and this one guessed with complete confidence.
That is a bigger concern than a blurred detail, because a customer might buy on the strength of what the back shows. We would not publish this clip for any item whose reverse differs from what the model guesses, and you cannot know that without checking each one.
Test 6: supplying the back of the bag
The fix costs nothing extra in generation fees, only an additional photo. Using image_with_roles, we supplied the front as first_frame and the rear still as last_frame:
{"model": "seedance-2.0-fast",
"prompt": "The backpack turns slowly on an invisible turntable from its front view to its rear view. Static camera. Background stays seamless light grey. No people, no added objects.",
"image_with_roles": [
{"url": "FRONT_PHOTO", "role": "first_frame"},
{"url": "REAR_PHOTO", "role": "last_frame"}],
"duration": 5, "resolution": "480p", "size": "adaptive", "seed": 5150, "generate_audio": false}
The ends are controlled. The in-between frames are still the model's own guess.
This was the best clip of the set. The opening matched the front photo at SSIM 0.95, the closing frame matched the rear at 0.95, and the tan straps are correct at the end. The clip cost $0.18 and rendered in 144 seconds, and mean colour error was 3.7, peaking at 5.6 late in the turn. At the side views around the midpoint the model still has to invent, so check the side pockets and strap attachments against the real item.
The workflow lesson is that photographs still matter, just fewer of them. Two good stills per SKU, front and rear, give you a controlled half-turn. A side still might tighten the middle, but a clip has only one first and one last frame, so that would mean chaining two clips, which we did not test.
The model page says the same request works for all three Seedance 2.0 variants, and the live per-second prices for each resolution and tier are on the Seedance 2.0 Fast API page. Our account was billed $0.0359 per second at 480p and $0.0770 at 720p, the low end of the range listed there, so check which tier applies to you.
Test 7: a hand opens the pocket
Lifestyle and in-use shots sell products, so we asked for something mechanical and specific:
A hand enters from the right, unfastens the brass buckle on the small front pocket and lifts the pocket flap to show the inside, then withdraws. Static camera. The backpack otherwise stays still. Background stays seamless light grey.
The hand looks believable. What it does to the bag does not make sense.
The hand itself was better than we expected: in the frames we sampled, the fingers looked natural and the arm entered and left as requested. The mechanics failed. The buckle is never undone. Around 2.8 seconds the whole flap lifts with the strap still fastened, and a small tab appears on the pocket body that is not in the photograph. Around 3.6 to 4.0 seconds the raised strap covers the brand patch, and OCR drops to fragments such as "TAR | ICK" in five of the 30 frames. At the end the flap stays in an odd rolled position.
A store selling this bag would be showing a closure that does not work the way the real one does. We would not use generated hands to demonstrate how anything opens, fastens, folds or attaches. Film those for real, on a phone if need be.
Test 8: getting the file ready for a product page
The delivered files are much heavier than a product page needs, and this part needs no API spend. We re-encoded three clips with ffmpeg and scored each result against the source with SSIM.
CRF numbers are not comparable across codecs, so read size and SSIM together.
H.264 at CRF 28 took the hand clip from 1,475 KB to 206 KB with an SSIM of 0.966, and the 720p tin from 2,146 KB to 279 KB at 0.978. CRF 32 went further, to 113 KB and 148 KB, at 0.952 and 0.969. AV1 at CRF 38 beat CRF 28 H.264 on both of those clips, but on the third clip we tested, the green bag arc, it produced 245 KB at 0.955 where H.264 CRF 32 managed 169 KB at 0.957. Mixed results on three clips do not justify a two-codec pipeline, so we would start with H.264 alone, using the +faststart flag.
Now do the sums for a category page. Twenty-four products at 206 KB each come to about 4.9 MB if every clip loads with the page. We would keep video off the grid by default, show the still image, and load the clip when a shopper opens the product or hovers over it. We did not run browser performance tests here, so that is a design recommendation and not a measurement.
Looping matters too, because a short product clip usually repeats. SSIM between the last and first frames was 0.62 for the green arc clip and 0.64 for navy, so both would visibly jump. The turntable and the hand clip scored 0.90 and 0.91 because they return to roughly where they began. An arc that travels one way should play once and stop on its last frame, or run forward and then in reverse. We did not test those two options on a live page.
Test 9: throughput, scoring and cost per SKU
We sent eight of the nine clips at the same moment, and all completed. Render times across all nine ranged from 93 to 279 seconds, with a median of 132. The 720p tin was the fastest of the set at 93 seconds, so a higher resolution did not mean a longer wait here. At a mean of 142 seconds and eight jobs at a time, 400 SKUs would take about two hours, assuming the rate holds, which we did not test beyond these eight.
Then we scored each clip as a store owner would, from sampled frames. This is our judgement, from one reviewer, and we built several tests to fail.
|
Clip |
Cost |
Render |
Verdict |
|
Green bag, text only |
$0.18 |
99 s |
Reject: different product |
|
Green bag, photo, arc |
$0.18 |
132 s |
Usable after trimming the first second or so |
|
Rust bag, arc |
$0.18 |
142 s |
Reject: wrong colour |
|
Navy bag, arc |
$0.18 |
279 s |
Re-render: pale pocket |
|
Tin, 480p |
$0.18 |
124 s |
Usable |
|
Tin, 720p |
$0.39 |
93 s |
Usable |
|
Bag, full turn, front only |
$0.18 |
148 s |
Reject: invented rear |
|
Bag, front and rear pinned |
$0.18 |
144 s |
Usable after checking the side view |
|
Hand opens pocket |
$0.18 |
118 s |
Reject: wrong mechanics |
|
Nine clips plus nine stills |
$1.85 |
Four usable |
|
The stills cost $0.027 of that total. The narration on the walkthrough video below is not included.
A catalogue budget, as an estimate: for 400 SKUs at the 480p tier, assume each needs 1.5 renders on average, which comes to 600 clips at $0.18, or about $108. Add 720p finals for your 40 best sellers at $0.39 each, about $16, and the total is around $124. The rate of 1.5 is our assumption and your rejection rate could be higher, since ours was five out of nine with hard tests included. The larger cost is review time. At one minute per clip, 600 clips take ten hours of somebody's attention, and that is the line a store owner should budget for most carefully.
What we did not cover: vertical and widescreen crops, native audio, and other product types such as a boot, a plaid blanket, a mug and a headlamp. Four of our stills were made for those tests. They were in our plan, but our connection failed while the jobs were being submitted, and we chose not to send them again without knowing whether the first attempts had reached the service. We have no results on them, so we make no claim either way.
Narrated walkthrough of the nine tests, with the generated clips.
A checklist for catalogue video
-
Always start from a photo of the actual item. Text-only clips passed our colour and text checks and still showed a different bag.
-
Supply every side you intend to show. Front and rear stills as first and last frames cost the same as a guess.
-
Use the first_frame role when the opening matters. A plain image input opened as low as SSIM 0.64.
-
Measure colour on every variant. The same prompt gave a mean Delta E of 1.6 on navy and 17.0 on rust.
-
Keep moving parts out of generated clips. Buckles, zips and closures should be filmed for real.
-
Run OCR across the frames, and overlay what must be exact. Large labels held up well. Fine print was not tested and should not be trusted to a generator.
-
Re-encode before upload and keep video off the grid. H.264 at CRF 28 cut our three test files by 81 to 87 percent.
-
Budget for review, and have a person watch every clip before it goes live. Our navy bag had the best colour score and still had a visible fault.
The bottom line
For a store catalogue, Seedance 2.0 Fast is useful within limits we can now describe. Given a photo and modest movement, it produced clean, readable and mostly faithful clips for under twenty cents each. Given two photos, it produced a half-turn that ends on the true back of the product. Asked to imagine anything it had not seen, or to demonstrate how something works, it produced confident fiction that a quick look might not catch. The rule we would give any store owner is simple: the model can move what you show it, but it should not be asked to tell customers what they cannot see. If you want to run the same tests on your own catalogue, we used APIMart for the generation requests, and the prompts above can be reused as they stand.