A product is harder than a face

Why the same techniques fail here

Everything in this block works better on people than on objects, and the reason is about the viewer rather than the model.

A face has natural variation. It changes with expression, angle and light, so a viewer’s tolerance for small differences is wide: they are used to the same person looking different in different photographs.

A product does not vary. It is a manufactured object with fixed proportions, a fixed label, a fixed logo. Any difference between two images is a defect, and the viewer can check it against the actual thing on a shelf.

So the tolerance that makes character LoRAs acceptable does not exist here.

What goes wrong specifically

The label. Text, which the repair block established cannot be generated reliably. A near-correct label is worse than none: it reads as a counterfeit.

Proportions. A bottle’s neck-to-body ratio drifts a few per cent between images. Individually imperceptible, obvious side by side, and immediately wrong to anybody who knows the product.

The logo. A trademark, which must be exact. An approximation is a legal and commercial problem, not an aesthetic one.

Material behaviour. How glass refracts, how a metallic finish catches light, how a matte surface falls off. Generated versions are plausible and not specific.

Colour. As the brief block covered, a specific colour is not something a diffusion model can hit.

The answer, and it is not a compromise

Photograph the product. Generate the environment. Composite.

The full method is in the previous block. Here is why it is specifically the right answer for consistency: a photograph of a product is identical every time you use it. Twelve images using the same cut-out are twelve images with exactly the same product, at zero risk. No LoRA, no drift, no label problem.

For a set, shoot the product once properly at several angles, build a small library of clean cut-outs, and reuse them across every frame. That library is an asset that serves this campaign and the next three.

When you must generate the product

A product that does not exist yet: a concept, a pitch, a packaging proposal. Then:

1. Model it, if you can. A simple shape in Blender takes an hour and gives you exact, identical geometry from any angle, with the label applied as a texture. This is the right answer for anything with a defined form, and it is free.
2. Train a LoRA on renders. If you have a 3D model, render 30 views and train. Now the product is in the model’s vocabulary and can appear in generated scenes with reasonable consistency.
3. Generate once, then reuse. Create the product image once, cut it out, and composite it everywhere: treating your own generated product exactly as you would a photographed one.

That third option is the one people miss, and it is usually the best. The moment you have one acceptable version of the object, stop generating it. Turn it into an asset.

The label, always

Whatever route you take, the label is set as type and applied afterwards. Generate or model the blank container, then place the artwork on it with a perspective transform, matched for lighting and curvature.

For a curved surface, the free path is a displacement map in GIMP: build a greyscale map of the curvature, apply it to the flat label layer, then multiply it over the product so the underlying shading shows through. Fifteen minutes, and it holds up at print size.

The exception worth knowing

Generated products are fine where the product is generic and not the subject of a claim: a coffee cup on a desk in a lifestyle image, an unbranded notebook, a plant pot. These are set dressing, and nobody will compare them against anything.

The line is whether the object is what the image is about. Set dressing can be generated. The hero product cannot, and pretending otherwise produces a specific, avoidable embarrassment: a client seeing their own product rendered with a label that nearly says the right thing.

Building the cut-out library properly

Since the photograph is the asset, shoot it once and shoot it well. An afternoon spent here serves every campaign afterwards.

For each product:

• Four to six angles: front, three-quarter left and right, a slight top-down, and one at the height the product usually sits on a table.
• Two lighting setups: one soft, one with a harder directional key, so you can match either kind of generated scene without reshooting.
• A clean background: mid-grey is better than white or black, because it contaminates the cut-out edge far less than either extreme.
• The product turned so the label is legible in at least half the frames.

Cut them out, save them as PNGs with transparency, and store them in a product/ folder alongside a short text file saying what the lighting was in each. That note is what lets you pick the right cut-out in eight seconds three months later rather than opening all twenty.

The reflection problem

One thing worth planning for. A product placed on a glossy generated surface needs a reflection, and the reflection has to come from your cut-out: a flipped, faded, slightly blurred copy on a layer beneath it, masked so it fades with distance.

If the surface is strongly reflective, the product also needs to reflect the scene, which the photograph obviously does not contain. For a matte product this is negligible. For chrome or glass it is the whole job, and it is the case where a 3D render beats a photograph, because a renderer can reflect an environment map you supply.