I spent ₹5,000 to learn that simple motion beats AI spectacle
I sell home décor on Flipkart—wooden coasters, ceramic planters, brass candle holders, that sort of thing. Average order value ₹800, mostly tier-2 cities, mostly gifting season and wedding orders. Last month I decided to test AI video for product ads instead of paying ₹2,500 per video to a local agency. I had eight products ready for a Diwali push, a ₹5,000 ad budget split between Flipkart Ads and Instagram, and access to three AI video tools on Kryotta: Luma Ray 2, Nova Reel, and Stable Image Core (for generating the base frames before animating them).
The plan was simple. Make eight videos, two per product category, run them as 15-second ads for five days, and track which ones got clicks versus which ones got skipped in the first three seconds. I'd read enough blog posts claiming AI video was "ready for ecommerce," so I wanted to see if that was true when you're spending your own money and competing against actual product photographers.
Four videos worked. Four were disasters. The difference wasn't the model—it was what I asked the model to do. Flipkart and Instagram users will sit through a slow 360° product spin with good lighting. They will not sit through a ceramic planter "emerging from a swirl of clay particles" or text that renders half-illegible because the model doesn't understand typography. Here's what I learned.
Luma Ray 2 handles simple camera moves, nothing else
I started with Luma because I'd seen demo reels that looked polished. My first prompt was for a set of brass coasters: "Camera orbits around a stack of four brass coasters on a white surface, slow rotation, studio lighting, 15 seconds." I uploaded a still image I'd made in Stable Image Core—clean product shot, soft shadow, nothing fancy.
Render time: 11 minutes. Result: smooth, usable, boring in a good way. The coasters stayed solid, the lighting didn't flicker, the motion was even. I uploaded it to Flipkart Ads as a sponsored product video. Over five days it got a 4.2% click-through rate on 1,847 impressions. Cost per click: ₹6.80. That's better than the static image I'd been running, which was hovering around 2.8%.
Then I tried to get clever. Same coasters, new prompt: "Brass coasters stack themselves one by one, then the camera pulls back to reveal the full set, elegant and smooth." Luma gave me 15 seconds of coasters that warped, duplicated, and at one point appeared to melt into the surface. The motion was there, but the object coherence wasn't. I didn't run it as an ad.
Luma works if you ask for camera movement and nothing else. Orbits, slow push-ins, gentle pans. The moment you ask the model to animate the product itself—stacking, pouring, unfolding—it falls apart. Render times stay under 15 minutes, which matters when you're testing multiple versions.
Nova Reel kept the product intact but added motion blur I didn't want
Nova Reel was faster—8 minutes for a 15-second clip—and better at keeping product details consistent. I tested it on a ceramic planter, same workflow: generated a base image in Stable Image Core, then prompted Nova with "Slow zoom into a white ceramic planter with green monstera leaves, the camera moves from wide shot to close-up, studio lighting, shallow depth of field."
The planter stayed a planter. The leaves didn't morph. The lighting was consistent. But every frame had a subtle motion blur, like the whole video had been shot at 1/30 shutter speed. It wasn't extreme, but it made the product look slightly soft, especially in the first three seconds when people decide whether to keep watching. I ran it anyway on Instagram as a Reel ad.
Result: 2.1% CTR on 2,203 impressions, ₹9.20 per click. Worse than the static carousel I'd been running. I checked the Instagram analytics—average watch time was 2.4 seconds. People were bailing before the zoom finished.
I tried a second planter video with a different prompt: "Static shot of ceramic planter on a wooden table, no camera movement, natural light from the left, 10 seconds." Render time dropped to 6 minutes. The motion blur disappeared. CTR jumped to 3.8% on the next batch of impressions, ₹7.10 per click. Not amazing, but usable.
Nova Reel is better than Luma at holding product shape, but the default motion settings add blur you don't need. If you're making ecommerce ads, explicitly tell it "no camera movement" or "static shot" and let the product itself be the focus. The algorithm will still treat it as video content, which gets better reach than a static image on Instagram.
Stability AI made the best base images, but animating them was a mistake
I'd been using Stable Image Core to generate the product shots before handing them to Luma or Nova. For a set of wooden coasters, I prompted: "Four round wooden coasters stacked on a white marble surface, top-down view, soft shadow, product photography, commercial lighting, sharp focus." Three iterations, picked the best one, 40 seconds total.
That image, uploaded as a static sponsored post on Flipkart, got a 3.6% CTR over five days. Then I tried to animate it in Luma with the prompt "Camera slowly rotates around the wooden coasters, smooth motion, 12 seconds." The coasters warped. The grain pattern shifted between frames. The shadow flickered. I scrapped it.
Stability AI is better at generating the hero frame than the video tools are at animating it. If you're on a budget, make one excellent still image and run it as a static ad. The CTR difference between a great still and a mediocre AI video isn't worth the extra render time and the risk of motion artifacts.
Text overlays don't work in any of these tools
I tried adding text to three videos: "Handmade in India" over the coasters, "Perfect for gifting" over the planter, "₹599 only" over a candle holder. I used text layers in the Kryotta editor before rendering, thinking the model would bake it in cleanly.
None of them rendered the text correctly. Luma turned "Handmade in India" into a blurry smudge that looked like a watermark. Nova Reel kept the text sharp for the first two seconds, then it started drifting and warping as the camera moved. The price tag "₹599" became "₹5?9" in one frame and disappeared entirely in another.
I ended up adding text as a separate overlay in CapCut after exporting the video from Kryotta. It took an extra ten minutes per video, but at least the text stayed readable. If you need text in your product video, add it in post. The AI video models don't understand typography well enough to keep it stable across 15 seconds of motion.
The winning formula: one product, one slow move, no cleverness
Out of eight videos, the four that actually drove clicks followed the same structure:
- One product in the frame, no props, no background distractions
- One camera movement: orbit, zoom, or static
- Studio lighting or natural light, explicitly named in the prompt
- 10–15 seconds, no longer
- Text added separately after rendering
The four videos that failed all tried to do something extra: multiple products in one shot, the product animating itself, "cinematic" lighting that looked muddy on a phone screen, or camera moves that were too fast and triggered motion blur.
The best-performing video overall was the brass coasters on Flipkart: 4.2% CTR, ₹6.80 CPC, made in Luma Ray 2 with the simplest possible prompt. The worst was a candle holder "rotating on a turntable with particles of light floating around it"—0.9% CTR on Instagram, and I killed the ad after two days.
Render times matter when you're testing multiple versions
Luma Ray 2: 10–15 minutes for a 15-second video
Nova Reel: 6–10 minutes for the same length
Stable Image Core: 30–50 seconds for a single hero frame
If you're testing five different products and two video variations each, that's 10 renders. At 12 minutes average, you're looking at two hours of render time. I ran all of mine in Kryotta's workspace overnight, queued them up before bed, and reviewed them in the morning. The team workspace feature let me share the final videos with my Flipkart account manager without downloading and re-uploading, which saved another 20 minutes.
Budget your render time the same way you budget your ad spend. If you've got ₹5,000 and five days to test, you don't have time to wait three hours for one video to render. Use the faster model (Nova Reel) for your first drafts, then re-render the winners in Luma if you need smoother motion.
When to skip AI video entirely and just shoot it yourself
AI video worked for me because my products are small, static, and simple to light. Coasters, planters, candle holders—objects that don't have complex textures or moving parts. If you're selling clothing, jewellery, or anything that needs to show fabric drape, clasp mechanisms, or how a product looks when someone's actually using it, AI video isn't there yet.
I tried making a video for a cotton table runner. The fabric texture looked like melted plastic in Luma, and the fringe detail disappeared entirely in Nova Reel. I ended up shooting it on my phone in five minutes: draped the runner over a table, filmed a slow pan, added text in CapCut. That video got a 5.1% CTR on Instagram, better than any of the AI versions.
Use AI video when the product is rigid, the motion is simple, and you're okay with "good enough." If you need "perfect," or if your product's selling point is texture and detail, spend the ₹2,500 and hire someone with a camera and a lightbox.
Questions people ask
Which AI video tool is fastest for product ads?
Nova Reel renders in 6–10 minutes for a 15-second video. Luma Ray 2 takes 10–15 minutes but handles camera movement more smoothly. If you're testing multiple versions quickly, start with Nova Reel.
Can I add text directly in the AI video, or do I need a separate editor?
Add it separately. Luma, Nova Reel and Stability all struggle with text overlays—letters warp, drift, or disappear across frames. Use CapCut or the Kryotta editor after rendering to add text that stays readable.
Do AI product videos perform better than static images on Flipkart and Instagram?
In my test, yes—4.2% CTR for AI video versus 2.8% for a static image on Flipkart, and 3.8% versus 2.9% on Instagram. But only when the video is simple: one product, one camera move, no effects. Overproduced AI videos performed worse than static images.
How much should I budget for testing AI video ads?
I spent ₹5,000 over five days testing eight videos across Flipkart and Instagram. That gave me enough data to see which styles worked. If you're just starting, ₹3,000 and three products is enough to learn whether AI video makes sense for your catalogue.
I'm still using AI video for new product launches, but I've stopped trying to make it look impressive. Simple spins, clean lighting, no motion effects. The goal isn't to wow anyone—it's to show the product clearly enough that someone clicks through to the listing. If you want to try the same workflow, the models I used are all available in one workspace at Kryotta, and you can queue up multiple renders without sitting around waiting for each one to finish.
