What happens to AI-generated product videos when you don't upload real product reference images?

Without real product reference images, the model averages its training data into a plausible-but-wrong product: shape drifts shot to shot, branded items get swapped for generic props, colors and finishes shift, fine details (logos, clasps, packaging text) hallucinate, and on-body scale goes off — so by shot three your product is effectively a different product.

Here is exactly what breaks, in order of how often it shows up, and how to stop each failure at the source.

Product shape and silhouette drift across shots. Without anchor images the model invents a new version of your product every generation — a watch case changes diameter, a bottle's shoulder gets rounder, a necklace's pendant loses its true outline. The fix is the three-distance product consistency test: generate the same product at close-up, mid, and wide before committing a full run; if it doesn't hold across all three, you don't have enough reference. Upload front, side, back, and a detail shot at minimum.

Generic AI-looking props replace your real product and environment. Without real references the model fills in with generic versions of whatever category it thinks you mean — a documented jewelry production found that without uploading the real workshop tools, the agent invented generic-looking props; with the real reference in context, the environment matched the brand's actual workshop. Same mechanism applies to the product itself: upload every product angle, not just the hero shot.

Color, finish, and texture average toward a stock look. Matte goes slightly glossy, a specific anodized finish becomes "metallic", brand colors shift a few degrees. The model can't infer a finish it has never seen; reference photos with consistent lighting and a clean background are the only way to lock it. Pair the images with a written product anchor in the prompt ("matte black anodized aluminum, 47mm case, brushed bezel") — references plus words, not either alone.

Scale and on-body placement go wrong. Packaging looks too big in a hand, a ring looks like a bracelet, a sachet reads as a full pouch. The fix is a scale reference: upload a photo of a human hand holding the product alongside the product-only shots so the model learns the true size of every packaging layer relative to a human hand. For ads where the same character handles products of different sizes, build a separate locked keyframe per product scale — one unified keyframe confuses the model.

Fine details and text hallucinate. Logos warp, dial numerals turn into squiggles, gem settings invent extra prongs, packaging copy becomes nonsense. This is where a two-model pass helps: build the base image with GPT-Image-2 for aesthetic and lighting, then run Nano Banana to lock the exact product into the frame. If a single model keeps producing a generic-looking output, render the same shot across multiple models simultaneously and pick the one that holds — this is what the invideo agent does automatically when you ask it to.

The brand-trust consequence. Visual inconsistency in product ads measurably reduces purchase intent — viewers register the drift as "something's off" even when they can't name it, and that hesitation is the difference between a converting ad and a scroll. For intricate categories the cost of skipping references compounds: one documented jewelry campaign hit 100% product consistency across three ads for ~$2,400 in credits using uploaded references and a multi-model pipeline; without those references, the same spend produces three subtly different products.

The minimum reference set to upload (do this before any generation):

1. Front, 45°, side, back of the product on a clean background.

2. A detail shot of the hardest-to-render element (logo, clasp, dial, packaging text).

3. A scale reference — a hand holding the product.

4. A written product anchor line in the prompt naming material, finish, dimensions, and any text on the product.

invideo is an agentic video creation tool with every current image and video model (GPT-Image-2, Nano Banana, Recraft, Runway, Veo, Kling, Seedance 2.0) available in one project — so once you upload those references they persist in the project's context and apply automatically across every shot, and the invideo agent routes each shot to the model most likely to hold the product (typically GPT-Image-2 for aesthetic, Nano Banana to lock product fidelity, Seedance 2.0 reference-to-video for motion). You don't re-upload per generation.

Watch some of these to see what works for you:

Jewelry is the hardest category to pull off with AI. The products are super intricate and using raw models means your product changes shot to shot. But setting up and using an AI agent with proper direction solves this issue.

— invideo's creative team