Features
AI Product Swap Video
By Ofrai Editorial (Use Case Research)
AI product swap video replaces the product in a shot you already have while keeping the camera move, lighting, and background locked. You shoot one clip per scene, then generate a variant per SKU from that same take — which turns a reshoot budget into a generation budget.

Chat and video use credits from your Ofrai balance.
Example prompts
Two people talking in a sunset greenhouse, plants and warm sky, cinematic mid-shot two-shot
One take, many product variants
The economics work because the expensive part of a product video is not the SKU — it is the shoot. Set, lighting, talent, and camera move stay fixed while the object in frame changes. A three-variant launch becomes one afternoon of capture plus a short generation queue, instead of three bookings.
What must stay locked versus what should move
Lock the camera path, the horizon line, the shadow direction, and the contact point where product meets surface. Let the swap absorb the object's own shape, specular highlights, and reflections. If the new product has a different footprint, re-render the contact shadow rather than letting the model improvise it — a floating object is the single most common giveaway.
Reference inputs that give the cleanest swap
Supply two references: the original clip frame and a clean product shot on a plain background. The clip tells the engine what the scene is doing; the product shot tells it what the object is. A lifestyle product photo with props baked in makes the model import the props too, which is almost never what you want. Use a cutout-grade image as the second reference.
Ratio, duration, and the loop-friendly cut
Most ad placements want 9:16 for feeds and 1:1 for marketplace cards. Render the swap at the same aspect ratio as the source clip so you are not cropping motion out of frame. For looping product demos, hold the first and last frame visually close — the model drifts most between frames, so anchoring both ends is what makes a seamless loop.
Motion clips that fail, and why
Fast handheld movement, heavy motion blur, and occluded products all break the swap: the engine has no clean read on the object silhouette. Slow push-ins, locked-off shots, and products with at least 60 percent of their outline visible swap reliably. When the source is hard, cut to a steadier moment in the clip rather than fighting it.
Which engine handles this job
Seedance 2.5 holds product geometry best on structured objects. Kling 3 is stronger when the scene has complex camera motion. Gemini Omni is useful when the swap involves reading scene context across frames. Wan 2.6 is a reasonable cost-first pass. Run the cheap engine for a look, then re-render the approved take on the strongest one.
Where this approach stops working
It does not cover genuine product demonstration — a hand applying the product, liquid pouring, material deforming under use. Those are real physical interactions the model will approximate badly, and a wrong pour reads worse than no video. Reserve swap workflows for hero shots, lifestyle placements, and variant changes where the product is held static in frame.
Product swap prompts written out in full
Same category, new SKU: "Replace the amber glass bottle on the shelf with a clear frosted bottle of the same shape and height, same cap colour family, matte label with no readable text, standing in exactly the same position and touching the same surface." Same product, new colourway: "Keep the bottle identical; change the label to deep navy with a thin gold rule, glass unchanged."
New size in the same line: "Swap the 30ml bottle for the 100ml bottle of the same design, standing side by side, both bases on the same surface line." The pattern is: state what stays, state what changes, state the physical relationship to the scene. Give the original product shot as the object reference and the source clip as the environment reference rather than describing the original in words, because written descriptions of packaging drift toward generic shapes within a couple of passes.
The settings that decide whether the swap holds
Check the aspect ratio, duration and resolution available for the selected video model. Aspect ratio must match the source clip exactly, because re-cropping a swapped object afterwards reintroduces the mismatch the swap was meant to solve. Duration should equal the usable length of the source clip: a five-second take asked to become ten seconds will stretch frames and the product will breathe. Generate above your delivery resolution whenever the label has to be legible, because text degrades faster than the rest of the frame.
Reading a failed swap frame by frame
Play a failed take back at quarter speed and find the first frame that goes wrong, because everything after it is downstream of that frame. If the product changes shape between frame 40 and 45, the failure is geometric and no prompt fixes it — re-render with a simpler reference or a different engine. If the product stays correct but floats, it is the contact shadow, and asking for the shadow explicitly usually resolves it in one pass.
If label text degrades progressively, the label occupies too much of the frame for the resolution, so generate larger or plan to composite the wordmark. If the background shifts slightly, both references are competing and the prompt needs to name which leads. If every single frame looks right but the motion feels wrong, the product was never connected to the surface — again a contact point, not a shape problem.
Credits across a SKU run
Model the run by SKU count rather than shot count, because that is how the budget behaves. A launch with three SKUs across two scenes is six outputs, and at a realistic two to four attempts per output you should expect roughly fifteen to twenty generations before you hold three keepers. The savings come from structure rather than from choosing cheaper engines everywhere.
Draft every SKU at low resolution on a cost-first engine to confirm the product reads correctly in frame, because that failure is compositional and cheap to detect. Produce the approved variants on whichever engine holds geometry best. And generate each SKU from the same source clip with the same prompt skeleton, changing only the product sentence, so that a fix on one variant teaches you something about all of them. Expect the first SKU in a run to cost two to three times the last.
Product swap versus object replacement in video
Both replace something in a clip you already have, and the practical difference is what the result has to survive. A product swap goes into a storefront, a marketplace card or a paid placement, so it is judged at thumbnail size beside competitors, with the label expected to be legible and the packaging expected to match the real product. That means clean cutout references, a matched aspect ratio and a colour check before publishing.
An object replacement inside a film or short is rarely inspected that closely — nobody verifies the stitching on a satchel in a lifestyle scene — so it tolerates a lower-fidelity result and a faster engine. If your deliverable is a listing, swap the product; if it is a scene, replace the prop. Confusing the two produces catalogue images with cinematic imperfection, which is the worst of both.
Who should skip product swap entirely
Skip it when the shot has to demonstrate the product working. A hand applying cream, liquid pouring, a hinge opening, fabric deforming under pressure — these are physical interactions, and a model that approximates them produces something a customer will notice is wrong. Skip it when the product has to be shown in an exact colour, because generated packaging shifts hue and a mismatch between the video and the product page becomes a returns problem.
Skip it when a specific person has to handle the product in a specific way, such as a child holding a toy, because hands are the weakest link in this pipeline. And skip it when you have one shoot, many SKUs and no appetite for art direction afterwards: the workflow works, but it delivers twelve variations that still need a human to choose between them, and that time has to be in the plan.
One last piece of process advice that matters more than any engine choice: keep the original un-swapped clips alongside every output. Campaigns get revised, legal teams ask for the source, and the next season reuses the same setup with a different product in the same scene. Storing the original takes costs almost nothing at the point of capture and is close to impossible to recreate later, because the shoot is the expensive part and the file is the cheap part.
Budget for that storage from the first day rather than after the first revision request arrives, because by then the shoot date, the set and the talent are all gone. Budget for it from day one.
FAQ
- Do I need the original footage to be high resolution?
- Higher resolution helps but is not the limiting factor. What matters is a clean read on the product silhouette and visible contact points. A sharp 1080p clip with the product clearly outlined swaps better than a 4K clip that is blurred or partly occluded.
- Can I swap between different product categories?
- Technically yes, but the scene has to stay plausible. Swapping a bottle for a different bottle works well. Swapping a bottle for a wristwatch will usually pull the surrounding scene with it, since reflections, shadows, and surface interaction no longer match. Match categories when the product stays in the same part of frame.
- How many variants can I generate from one clip?
- There is no hard cap in the studio, but quality degrades if you push too many substitutions off the same source frame. Five to eight variants from one take is a reliable working range. For a larger catalogue, shoot separate clips per product family so the engine is only changing the SKU, not the scene.
- Will the swapped product keep its logo and label text?
- Usually yes when you supply a clean product reference and the label occupies a reasonable part of frame. Logos on curved or low-contrast surfaces are the usual failure. Generate at a larger resolution than you need and downscale, because text legibility degrades faster than the rest of the frame.
- Can I use the result commercially?
- Commercial rights depend on the model you generate with and your Ofrai plan, not on this workflow. Check the specific model page and your subscription terms before publishing paid media. Keep your source clip too — you will want the un-swapped original if a campaign is revised.
Related generators
- AI Object Replacement Video
- AI Photo Dance Video Generator
- AI Virtual Try-On Video
- AI Dance Transfer For Choreography Reuse
- AI Motion Transfer Video Generator
- AI Outfit Swap Video
- Seedance 1.5 Pro on Ofrai
- Photo to video
- AI Animal Video Generator
- AI Anime Video Generator
- AI Furry Video Generator
- AI Realistic Video Generator