Lore

AI product photography

Put Your Product EXACTLY Where You Want with AI | Weavy & Nano Banana

The video argues that AI image models have solved realism but not control, and that control comes from turning a product photo into a deliberately crude 3D 'mannequin'. The mannequin fixes the product's pose, camera angle and position, while the image model (Nano Banana in Weavy) supplies realism and detail from separate references.

Ľudovít Nastišin · 2026-02-12 · English

Key ideas

  1. The video states the problem: realism is no longer the difficulty for AI, but precise control is. Examples are showing a specific part of the label, or deciding where two products touch, how close they sit and which overlaps.

  2. Its proposed answer is to split the job. A rough 3D stand-in decides where the product is and how it is oriented, and the image model supplies how it looks. The author says the ugly 3D version is fine because 'I am not using it for realism.'

  3. The author wants one reusable scaffold: a long fixed prompt, annotated reference images and a labelled node layout that he says you don't touch. Per image you supply only the product image and a scene description, so the long prompt is 'not so scary, it's just long.'

  4. Placement replaces prose for spatial relationships. Several posed mannequins can be combined into one layout reference, so compositions that would otherwise need a long, expert text description can be set up visually.

  5. Output quality still depends on the inputs. The author says more context for the model, such as extra angles and close-ups, and more effort in the scene description both give better results. He openly shows imperfect outputs: some contact-sheet panels are wrong, the size is guesstimated when aspect ratios differ, and the Antarctica scene is not logical. He argues the angle and placement logic still holds.

  6. He presents the results as raw material for banners or short animated reels, and says the workflow is free, easy to use and shared in cleaned-up form via the description.

  7. Garbage 3D product mannequin — Turning a single product image into a crude, textured 3D model via a 3D node, to be used only to set the product's pose, camera angle and position, not realism. Apply: Plug the product image into a 3D node (Rodin v2 or Meshy, any that gives a texture), then rotate, orbit the camera and position it until the desired part of the label shows. This is then applied to the final image.

  8. Nine-panel context sheet — An extra reference generated from the single product image that shows the product from different angles and close-ups, giving the model more context about how it looks. Apply: Feed the product image into a pre-made template with Nano Banana asking for a nine-panel view with at least three close-ups and different rotations, then use the result as the supporting image C.

  9. Compositor annotation layers — Compositor nodes that work like simple Photoshop layers, combining a background image (such as the 3D node export) with a text layer positioned anywhere, so each reference is labelled. Apply: Use the 3D export as the background and add a text node reading 'image A'. Do the same for the product detail image (B, can be small) and the additional details image (C), repositioning the text via Edit.

  10. Master prompt with reference-image roles — A long, fixed prompt that tells the model which inputs it receives and what each does: image A controls pose and camera only, image B is the detail and texture reference, image C is supporting material. Apply: Leave the scaffolding untouched and only enter your product and a scene description; the prompt in the middle updates itself, and you feed it to the Nano Banana node with the 3D node as the first reference, product detail second and extra details third.

  11. Switchable scene-description notes — Several prepared scene descriptions (for example product on a shelf, or in a shopping bag) kept as separate notes that drive the same workflow. Apply: Either rewrite the scene text directly or prepare multiple notes and connect the one you want, so the master prompt updates automatically.

  12. Multi-mannequin composition — Copying the 3D node several times to place multiple products, or different products, in one layout reference, controlling how they touch, how close they sit and which overlaps. Apply: Duplicate the 3D node (three toothpaste tubes, or a second bottle model), position and rotate each, merge them in a compositor into image A, and add the annotation step; the rest of the workflow stays the same.

  13. Batch runs and pick the best — Generating several variations at once instead of re-running a single generation until it works. Apply: Add four runs, hit 'run selected', and pick the best version.

Insights

The 3D asset's geometry quality and appearance don't matter, but it must carry some texture, even a bad one. Without a texture the model estimates the product's position in space poorly. Rodin v2 and Meshy are described as interchangeable as long as they output a texture.

With a 9:16 output and 1:1 reference images, the model has to guesstimate the product's size and fill pixels above and below. The rotation still lands in line.

A single product image is enough, but the author says it helps to generate extra context (a nine-panel sheet of angles and close-ups) because the model has more to work with. He didn't specify what each panel should contain, only that there should be at least three close-ups plus different rotations.

The third reference (supporting material) is described as rarely used by the model, yet the author feeds it in anyway: 'You never know which part will it actually use.'

Reference images are labelled with text layers in the corner, so the prompt can refer to 'image A/B/C' and assign each a distinct role. The layout image is told to control pose and camera exclusively and to ignore materials and lighting. The detail image is told to carry textures and everything that makes the product the product.

Dragging or rotating the 3D node updates the layout reference image, and re-running the model then updates the product's view angle. Orientation is edited visually, not in words.

For multi-product scenes the author notes that the downstream workflow stays the same, because the compositor merges everything into one layout image. The same pipeline handles one, three or two different products.

Instead of re-running a single generation until it works, the author prefers to queue four runs and pick the best.

Better scene wording improves results. The author suggests changing 'ice cube' to a piece of ice, specifying depth of field and adding a polar bear, and says the more effort in the description, the better the result.

In the final demo the second, differently posed bottle was added and the prompt asked for the bottles leaning on each other. The author reports the polar bear and the touching bottles appeared, although he isn't sure of the phrasing.

«Realism isn't much of a problem for AI anymore, but control is still a challenge.»

— 00:00

«So how I deal with that? I take the product image and turn it into garbage 3D version.»

— 00:06

«Looks terrible, I know, but that's fine. I am not using it for realism.»

— 00:13

«I am using it as a product mannequin.»

— 00:17

«part is not so scary, it's just long.»

— 03:34

«And you basically don't touch this part, you don't touch this part,»

— 04:57

«You never know which part will it actually use.»

— 04:51

«However, it needs to be the one that can give you also a texture.»

— 10:46

«It really doesn't matter that it's a garbage texture, but it needs to be there.»

— 10:51

«Otherwise, the model won't do a great job in estimating where in the space the product actually is.»

— 10:57

«The more effort you will give to this description, the better result there will be.»

— 15:24

«So you can see it actually works. I did not lie to you. there is a polar bear»

— 17:32

Reception

Overwhelmingly positive and grateful, with viewers praising the tutorial as practical and useful; the few critiques (distracting background music, a missing workflow link, 3D navigation difficulty) are minor and mostly answered helpfully by the creator.

A concrete, reproducible approach to compositional control: a throwaway 3D model sets pose and layout, and a labelled multi-reference prompt sets appearance. The demos are admittedly rough, but the idea that placement, not text, should carry spatial instructions is the transferable part.

17:44

↳ Ľudovít Nastišin · YouTube

Watch original