How do you test AI-generated video for product consistency across close-up, mid, and wide shots?

Before committing a full run, generate the same product in three shots from one locked reference — a close-up, a mid, and a wide — and verify the product holds in all three. If it drifts on any of them, fix the reference stack before you spend more credits. If it holds, you're cleared to scale the run.

Run the three-distance test like this:

1. Lock the product reference stack first. Upload every angle of the product (front, side, back, top, close-up of detail) plus one scale reference — a hand holding the product so the model learns its true size relative to a human. For intricate products like jewelry, also upload a clean detail crop. This is what the close-up, mid, and wide shots will each be anchored to.

2. Generate one close-up, one mid, and one wide from the same anchor. Use a single locked keyframe image as the visual anchor and ask the invideo agent to render the product at all three focal distances against the scene you actually plan to shoot. The invideo agent is an agentic video tool with every current image and video model available inside it, so the same reference flows through whichever model is rendering each distance.

3. Apply a three-question check to each rendered shot. For every one of the three shots, verify: does the product's shape and silhouette match the reference, does the color/material/logo read correctly, and does the product sit plausibly in its environment at that scale. If a close-up reads right but the wide loses the logo, or the mid renders a different shade, your reference stack is incomplete — not your prompt.

4. If a distance fails, fix at the source — don't iterate on prompts. Two fixes carry: add the missing reference angle (the wide failing usually means no scale-reference image; the close-up failing usually means no detail crop), or render that one shot across multiple models simultaneously and pick the one that holds. For jewelry-grade detail, a two-model pipeline works — build the base image with GPT-Image-2 for aesthetic, then run Nano Banana to lock the exact product into it.

5. Only after all three distances hold, commit the full run. This is the gate. As Hridaye, invideo's creative director, puts it: "Generate a close, a mid and a wide — with the product in every one. Confirm the piece holds before you commit the run." In one documented jewelry campaign, this gate kept product consistency at 100% across three finished ads (~$2,400 total) — the test caught drift early enough that no full batch had to be rerun.

6. Keep validating during the run. As clips lock, pull them into your editor and assemble in parallel — drift that survives the gate usually surfaces when shots sit next to each other. If you see it, regenerate that single shot against the locked reference rather than restarting the batch.

Watch some of these to see what works for you:

Generate a close, a mid and a wide - with the product in every one. Confirm the piece holds before you commit the run.

— Hridaye, invideo's creative director