3_OG — Hifromandrew

optin 1 - A 3D blockout doesn't help the model imagine the space. It stops it from imagining, which is the only way to get the same room twice. opton 2 - Unlocking spatial coherence across more than one image.

I found out very quickly when I started generating images in late 2025 that I am not very good at describing a scene in the detail I required. Can anyone?

That sounds like a prompting problem. It isn't. You can write a paragraph about a kitchen and get something good back. Write the same paragraph again and you get a different kitchen. The description was never the constraint — the model was filling everything I hadn't specified, and I couldn't specify enough.

From gathered experience with pure text prompting, I learnt that certain models at the time were simply not able to hold coherence for anything larger than simple inpainting results. Generations often led to scene manipulation, which meant models were too small in size and the training data pipelines were not yet mature enough for larger models to reason about complex requests. The model would generate with little to no constraining from my inputs.

Realistically, after testing models before and after, there is only one model on the market currently that can maintain scene geometry when given a 3D blockout as a strict scene and geometry lock. Basically skinning — a skin without changing it. That is Nano Banana Pro.

3D · Blender
AI Image Model · Nano Banana

Placeholder note — what this group demonstrates, running wider than the stamp.

Placeholder body copy. Set the problem, state the constraint, name what was tested. Prose only, calm and readable — this is the substance a crawler indexes and a reader came for.

2D · Photoshop
AI Image Model · Nano Banana Pro

Placeholder note — first test after completing the base image and the initial look set.

Placeholder body copy. Set the problem, state the constraint, name what was tested. Prose only, calm and readable — this is the substance a crawler indexes and a reader came for.

3D · Blender

Placeholder note — progression of the 3D model.
1. Placeholder step one.
2. Placeholder step two.
3. Placeholder step three.

Placeholder note — examples of generation at each stage of 3D blockout complexity.

Placeholder note — examples of responses given to the model.

Placeholder body copy. Set the problem, state the constraint, name what was tested. Prose only, calm and readable — this is the substance a crawler indexes and a reader came for.

Placeholder note — the earliest blockout and the generations it made.

Placeholder note — months after, revisiting the interior.

must use the provided image @<id> as the first frame.
static tripod, surveillance-style framing.
edge-pixel lock condition maintained across the loop.

Closing placeholder — the takeaway a business can actually use.

Driving AI models with 3D blockouts