A Video From One Product Photo Cost $4.25 Instead of $2.42 Because of Seedance 2.0 Defects
A marketplace seller needs a 15-second vertical video, but has only one photo of a bottle and no budget for a shoot. The author built a pipeline: Gemini draws a six-frame storyboard, a text model writes the prompt, and Seedance 2.0 shoots the video. The first take costs $2.42, but it contains four types of defects, such as a slogan with the letter «Е» instead of «Э». Three reshoots added $1.83, and the total grew to $4.25, so budget almost double for a video.
AI-processed from Habr AI; edited by Hamidun News
Marketplace sellers often ask for the same thing: they have a single product photo, they need a 15-second vertical ad, and they have no money for a physical shoot. The author of a write-up on Habr built a pipeline that produces such a video through an API and calculated what it costs in practice. The first take, meaning the first generated version, costs $2.
42 (images, prompt and Seedance 2.0 video). But the video had defects: for example, in the slogan the model drew «Е» instead of the letter «Э».
The fixes (three reshoots of broken shots) added another $1.83, and the total bill came to about $4.25.
If you make such videos for clients or build them into a service for sellers, count on a cost almost double the price of the first take.
How the pipeline works
The scheme has five steps. First, the product photo and a brief go to an image generator. In the test, the photo of a fictional serum, «СЕВЕР», was processed with gemini-3.1-flash-image. The generator draws a storyboard: six frames in a 2-by-3 grid, with timing under each one. Then the storyboard goes to a text model, which writes a video prompt based on the frames, with timecodes. The third step is generation with bytedance/seedance-2.0 through OpenRouter: 15 seconds, 9:16 format, 720p, with sound. Steps four and five are a frame-by-frame review and reshooting the broken shots one at a time. Text is added later, in editing.
OpenRouter serves video models asynchronously. You create a job with POST /api/v1/videos, then poll GET /api/v1/videos/{id} until the status becomes completed. The finished response contains a link to the file and usage.cost in dollars, so you don't need to calculate the price of a take from tokens.
Seedance has two main modes. In input_references mode, images serve as guidance: the model takes the packaging and colors from the photo and the order of shots from the storyboard, but builds the frames itself. In frame_images mode with first_frame, the image becomes the exact first frame.
The first mode is for the whole video, the second for reshooting a single shot. There is also a third option: first_frame plus last_frame, where the model fills in the motion between two frames. To keep the packaging from being redrawn, the author names the role of each image at the start of the prompt: the first is the only reference for the bottle, do not change the label; the second is only for the order of shots.
For a reshoot, he uses ffmpeg to cut out the last frame of the previous shot, passes it as first_frame and asks the model to continue the motion. The join comes out without a jump. Sound is turned off in the inserts: the price did not change (about $0.15 per second with or without sound), but one reshoot with sound failed moderation, which judged the generated music to resemble someone else's.
Four types of defects
The first take looks finished, but a frame-by-frame review finds four problems.
The first is an error in the on-screen text. The slogan came from the storyboard, the text model carried it into the prompt, and the video model drew it with a typo. Now the video prompt must contain no new text except the label inscription. The author tells the text model directly: "Do not add slogans, titles or any new text."
The second is a "glued-on" label. During a camera orbit the background changes, but the label keeps facing the lens. From a single photo, the model knows only one side of the bottle. A lateral tracking shot or an orbit of no more than a quarter circle is safer. Several angles as input would probably help, but the author did not test this.
The third is a jump at the join. Two adjacent shots are filmed from almost the same point, and the bottle shifts by a couple of centimeters. Adjacent shots should differ noticeably in scale or angle.
The fourth is ignored transitions. Instead of a soft dissolve between the fourth and fifth shots, the result was a hard cut, and the camera turned the wrong way at the join. The continuation should be described through visible motion: "the whole scene slides to the right" is more reliable than "the camera moves left." It is better to do transitions in editing.
What it costs and what the limitations are
The storyboard and product photo (2 images at 1K) cost $0.14, the prompt from gemini-3.5-flash about $0.01, and the 15-second video with sound $2.27. Together that is $2.42. Three reshoots of 4 seconds each without sound cost $1.83, for a total of about $4.25. 27 seconds of video were paid for and 15 made it into the final video, so the spend ratio is 1.8. Prices are as of October 8, 2026, excluding fees.
Steps 1–4 take about half an hour, of which eight minutes are spent waiting for the video. A 4-second shot takes about three minutes to generate. Step 5 eats the rest of the evening, and the author could not automate it fully.
The API also has limitations. Seedance 2.5 through OpenRouter rejected photorealistic images of people with the error InputImageSensitiveContentDetected.
PrivacyInformation. The author did not test this on version 2.0, but the model line and ByteDance's protection are shared.
So a video of a product in the hands of a hero, using your own character, can't be made this way, and you have to describe the hero in words. Rejected requests cost nothing. Moderation also triggers after generation: in the tests, both video and music were blocked.
Charges applied only to successful attempts, but the waiting time was lost. In addition, the model keeps the inscription from the photo but likes to tweak it, so every inscription must be checked against the original letter by letter in every shot.
Before delivery, the author checks six items: inscriptions match the original; there is no new text in the frame; on an orbit the label turns together with the product; adjacent shots differ in scale or angle; motion at the join goes in one direction; and there is no accidental speech or singing in the audio. The last item is easy to automate: the whole clip goes to gemini-flash-latest (video inline, up to ~18 MB), and the model lists all speech with timecodes. The first five items are still checked only by eye, using a contact sheet of frames.
The main takeaway: a video from a single photo comes together in about half an hour, but the first take is almost never final. Budget as much for fixes as for the first run, keep text out of generation, and check the label frame by frame.
- Nov 18, 2026Alibaba cut Qwen4-Max's cost per task by 67%, but GPT-6.1 Sol is still cheaper: $3.27 vs. $1.27
- Nov 18, 2026Meituan open-sources 1.6-trillion-parameter agentic model trained without Nvidia
- Nov 12, 2026Google opens preorders for Gemini glasses starting at $395 and blocks covert recording
- Nov 12, 2026Samsung Is Preparing a Paid Galaxy AI Subscription: From $4.99 a Month and Credits Instead of Unlimited Use
Key Questions Answered
How the pipeline works?
The scheme has five steps. First, the product photo and a brief go to an image generator. In the test, the photo of a fictional serum, «СЕВЕР», was processed with gemini-3.1-flash-image. The generator draws a storyboard: six frames in a 2-by-3 grid, with timing under each one. Then the storyboard goes to a text model, which writes a video prompt based on the frames, with timecodes. The third step is generation with bytedance/seedance-2.0 through OpenRouter: 15 seconds, 9:16 format, 720p, with sound.
What are the key details on Four types of defects?
The first take looks finished, but a frame-by-frame review finds four problems. The first is an error in the on-screen text. The slogan came from the storyboard, the text model carried it into the prompt, and the video model drew it with a typo. Now the video prompt must contain no new text except the label inscription. The author tells the text model directly: "Do not add slogans, titles or any new text." The second is a "glued-on" label.
What it costs and what the limitations are?
The storyboard and product photo (2 images at 1K) cost $0.14, the prompt from gemini-3.5-flash about $0.01, and the 15-second video with sound $2.27. Together that is $2.42. Three reshoots of 4 seconds each without sound cost $1.83, for a total of about $4.25. 27 seconds of video were paid for and 15 made it into the final video, so the spend ratio is 1.8. Prices are as of October 8, 2026, excluding fees. Steps 1–4 take about half an hour, of which eight minutes are spent waiting for the video.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.