I wanted twelve Reels ready before the Diwali rush. I got seven usable ones.
I run a small homeware brand—hand-poured candles, brass diyas, ceramic planters—and sell through Instagram and a Shopify store. Last October, I decided to make a dozen product Reels using Luma Ray 2 and Nova Reel instead of shooting everything with my phone. The goal was simple: have enough video content banked before Diwali orders started flooding in, so I could focus on packing and shipping instead of scrambling for fresh posts.
I spent ₹2,400 on credits across both platforms and four evenings testing prompts. Twelve videos later, I had seven I'd actually post, three that looked passable if I squinted, and two that belonged in a horror film. Here's what worked, what didn't, and the exact prompts I used.
The ones that looked real: single product, simple movement, neutral background
My best result was a brass diya rotating slowly on a white marble surface. I uploaded a photo I'd taken with my phone—decent lighting, plain background, no shadows eating half the product—and gave Luma this prompt: "Smooth 360-degree clockwise rotation on marble counter, soft natural light from the left, shallow depth of field, product stays centered." Six seconds, 1080×1920 vertical. It rendered in forty seconds and looked like I'd hired someone with a turntable rig.
The second winner was a close-up push on a candle label. Nova Reel handled this one. Prompt: "Camera slowly pushes in on product label, background gently blurs, warm afternoon light, no movement in the product itself." I used a photo where the candle was already centered and the label was sharp. The AI kept everything stable—no weird warping, no label text turning into gibberish halfway through the zoom.
What these two had in common: one product, one type of movement, and a background that didn't compete for attention. The AI models are very good at rotating or zooming on a single object when the input photo is clean. They're less good at inventing complex scenes or handling multiple elements.
I made five more videos in this vein—a planter rotating, a diya set with a slow pan across, a candle with a top-down spiral—and all five were usable. Total time: about ninety minutes, including the renders and two re-dos where I tweaked the lighting description.
The ones that looked fake: hands, multiple products, anything "dynamic"
I tried to make a video of a hand placing a diya on a table. The prompt was: "Hand gently sets brass diya onto wooden surface, fingers release and move away, natural motion." What I got back was a hand with six fingers that melted into the table like warm butter. The diya teleported three centimeters to the left mid-video. I tried this twice with Luma and once with Nova. All three attempts failed in different creative ways.
AI video models cannot reliably generate human hands in 2025. They especially can't animate them. If your product needs to be held, shown in use, or handed to someone, shoot it with your phone. It'll take you five minutes and look a thousand times better than the Cronenberg nightmare the AI will give you.
The second disaster: I wanted a video showing three candles arranged in a triangle, with the camera circling them. Prompt: "Slow orbit around three candles on a tray, camera moves smoothly, candles stay in focus." The result had two candles for the first three seconds, then a fourth one materialized, then one of the original two slid sideways like it was on ice. The tray changed color halfway through.
Multiple products in one frame confuse both models. They can handle a "set" if it's static—I got a decent shot of five diyas arranged in a line by asking for a simple pan—but the moment you ask the camera to move around multiple objects, the spatial relationships fall apart. One candle will drift, another will change size, and the shadows will point in three directions.
Backgrounds matter more than I expected
I tried making a video with a "festive Diwali background—marigold flowers, rangoli patterns, warm bokeh lights." Luma gave me a blurry orange-and-yellow mess that looked like someone had smeared turmeric on the lens. The product—a ceramic planter—was sharp, but the background was a hallucination of what the model thought "festive" meant. It wasn't unusable, but it didn't look like a real place.
The best backgrounds are the ones you photograph yourself. I shot a corner of my workspace with some fairy lights and a white wall, then placed products there and took reference photos. When I fed those into the AI with prompts like "gentle zoom on product, background stays soft and out of focus," the results looked natural because the background already existed. The AI just had to animate the camera move.
If you don't have a good background, stick to white, marble, wood or black. Prompts like "clean white surface with soft shadows" or "dark wooden table, moody lighting" work because the models have seen a million product photos on those surfaces. Asking for "vibrant festive Indian décor" gets you a fever dream.
The prompts that worked vs. the ones that didn't
Here are three prompt pairs—what I tried first, and what actually worked.
Failed: "Dynamic 360 spin with the product rising slightly, colorful background, energetic feel."
Worked: "Slow clockwise rotation on white surface, product centered, soft shadows, 6 seconds."
Failed: "Camera swoops around the product like a drone shot, fast and exciting."
Worked: "Gentle camera arc from left to right, product stays in frame, shallow depth of field."
Failed: "Show the candle being lit, flame flickers, warm glow spreads."
Worked: "Static shot of lit candle, flame stays steady, warm light on the product, slight camera push-in."
The pattern: specific, slow, simple. The moment I used words like "dynamic," "swoops," or "energetic," the model tried to do too much and the physics broke. When I described one movement, one light source, and one focal point, I got results I could post.
What I'd do differently next Diwali
I'd shoot better reference photos. Half my re-dos happened because the input image had uneven lighting or a cluttered background, and the AI amplified those problems. A clean photo with one light source and no distractions gives you a much better starting point.
I'd also batch the work. I tried to make one video, review it, adjust, make another. That took forever. The better approach: pick five products, write five simple prompts, render them all, then review the batch and tweak the ones that need it. Luma and Nova both let you queue requests; I just didn't use that feature until day three.
And I wouldn't bother with "lifestyle" shots. The AI can't put your product in a real kitchen or show someone using it at a dinner table. Those need a real camera and a real person. Save the AI for the product-only shots—spins, zooms, close-ups—and shoot the context yourself.
The cost vs. shooting it myself
I spent ₹2,400 on credits and got seven usable Reels. If I'd shot those same seven with my phone, I'd have spent maybe ₹500 on props (a marble tile, a wooden board, some flowers) and three or four hours setting up, shooting, and editing in CapCut. The AI saved me time, not money.
But the time savings mattered. I banked those seven videos in early October, then spent Diwali week packing orders and replying to WhatsApp messages instead of trying to shoot content between shipments. That's the real value: you can make the videos when you have time, not when you need them.
If you're a solo seller juggling product creation, customer service, and social media, that shift is worth the cost. If you've got a team or you enjoy shooting video, the AI probably won't save you anything.
What this means for festive season prep in 2025
AI video isn't going to replace your phone for every type of content. It's not going to show your product being unboxed, worn, or used. It's not going to film your workspace or your hands assembling something. But for the simple product showcase—the spin, the zoom, the slow pan—it's faster than setting up a tripod and cheaper than hiring someone.
I'll use it again next Diwali, but I'll also keep my phone charged. The best Reels I posted during the season were a mix: AI-generated product spins cut together with phone footage of the packing process and a few customer video testimonials. The AI gave me the polished product shots I didn't have time to shoot; the phone gave me the human moments that made people trust the brand.
If you're prepping for wedding season, Holi, or next year's festive peaks, try making three or four test videos now. Use simple prompts, clean photos, and neutral backgrounds. See which ones you'd actually post. Then you'll know whether it's worth the cost when the orders start rolling in.
Questions people ask
Can I use AI video for ads, or will Meta reject them?
Meta doesn't reject ads because they're AI-generated. They reject ads for policy violations—misleading claims, prohibited content, low-quality visuals. I've run Luma-generated Reels as ads without issues. Just make sure the product is clearly visible and the video doesn't look glitchy.
Which platform is better, Luma or Nova?
Luma Ray 2 handles rotations and orbits better. Nova Reel is faster and slightly cheaper, and it's better at close-up zooms. I use both depending on the shot. You can try them side-by-side in Kryotta's video workspace without committing to one platform.
How long does each video take to render?
Luma usually takes 30–90 seconds for a six-second clip. Nova is faster, often under a minute. If the queue is busy, add another minute or two. It's not instant, but it's faster than setting up a product shoot.
Do I need to disclose that the video is AI-generated?
There's no legal requirement in India as of now, but Instagram's policies encourage transparency. I don't add "AI-generated" to every caption, but if someone asks, I'm honest. Most customers don't care how the video was made—they care whether the product looks good and arrives on time.
I made those twelve videos because I didn't want to spend Diwali week shooting Reels between packing orders. Seven of them worked, and that was enough. If you're planning for the next festive rush, try a few test renders now—before the pressure's on. You can experiment with both Luma and Nova in one workspace at Kryotta and see which prompts give you results worth posting.
