3D & Blender · Fundamentals

What to elaborate in Blender — and what not

A practical checklist for the RAY-L workflow.

B1 and B2 explained the principle: Blender provides the deterministic structure, the AI handles the stochastic elaboration. But with the first real project comes a very practical question: what exactly do I build in Blender now — and what do I deliberately leave out?

This checklist answers that. It's distilled from B1, B2, and the Highland Cottage case study — as a direct decision aid for building your own scenes.

This belongs in Blender

Camera — position, height, focal length, angle of view. The single most important decision. It defines the composition and can't be corrected after the fact.

Basic geometry and proportions. Rooms, buildings, objects — at their actual scale and in the right spatial relationships to one another. ControlNet Canny extracts the edge structure from this.

Object placement. Exactly where the product, the piece of furniture, the figure stands. This position persists through every subsequent step.

Structural textures. Not for visual elaboration — but so ControlNet has something to detect. A tile pattern, a grout grid, a wood grain with its direction of lay. Surfaces without any texture give Canny nothing to work with.

Brand-identity-critical elements. Logos, typography, exact brand colors — anything that must not be left to AI interpretation.

Rough light direction. A single light source that roughly indicates where light and shadow should fall in the image. This indirectly influences what ControlNet reads as structure.

This doesn't belong in Blender

Fully elaborated materials. PBR shaders with roughness, reflectance, subsurface scattering — unnecessary effort for the RAY-L workflow. The AI supplies the material feel from its training.

Detailed lighting setups. Multiple light sources, precise intensities, color temperatures — all of this is interpreted by the AI from the prompt. A complete lighting setup costs time without showing up in the result.

Environmental details and props. Plants, props, background objects that aren't compositionally relevant. The case study shows: an entire cottage garden emerged without a single plant model in Blender.

Final renderings. No path tracing, no long render times, no post-processing. The Blender result is a tool for ControlNet — not a finished image.

Atmospheric effects. Fog, rain, light scattering, bokeh — anything that creates mood rather than structure. That's exactly where the AI is strongest.

This is optional — depending on the project

Material cues for specific elements. When a material must be precisely defined — a particular parquet pattern, a branded fabric — it doesn't go fully into Blender, but as a reference image into the AI workflow.

Multiple cameras. Useful when different perspectives of the same scene should be tested before the first RAY-L pass starts.

Placeholder geometry for figures or people. If a person should appear in the image, ControlNet needs a rough figure — proportions and posture are enough, details are handled by the AI.

Practical sequence

  1. 01Set the camera. Before anything else. Test multiple variants if needed.
  2. 02Build the basic geometry. Rooms, objects, proportions — at final scale and position.
  3. 03Add structural textures. Wherever ControlNet needs information that can't be derived from geometry alone.
  4. 04Set a rough light source. One direction, no elaboration.
  5. 05Canny export and ControlNet pass. Only now does the AI enter the picture.
  6. 06Prepare reference images — if specific materials or a target mood need to be defined.

The short version

When you're unsure about an element, one question helps:

Would this element change if I repeated the same project tomorrow with different materials or a different mood?

If no — it belongs in Blender. It's part of the structure.

If yes — it belongs in the prompt or as a reference image. It's part of the interpretation.

That one question replaces most detail decisions.

One exception: mask-based outpainting

This checklist applies to the ControlNet Canny workflow — the entire scene is reinterpreted, with geometry providing only the structure.

For mask-based outpainting (planned for RAY-L v1.1+), the logic is partially reversed. When a product or packaging needs to be embedded in the scene as a fixed, unchangeable element, that exact element must be fully elaborated — with materials, textures, and lighting. A mask protects it from AI interpretation. Only the surroundings outside the mask are generated by the AI.

So the logic flips: not "what do I hand to the AI," but "what do I protect from it." Both workflows follow the same underlying principle — deliberate decisions about where control ends and interpretation begins — just by different means.