Odicon AI: Constrained Generative Placement for Interior Design

Precision is the interface

Interior design renderings fail when they look good at a glance and wrong in practice. Odicon AI is built around one principle: no free-form imagination.

It only generates what the scene, the catalog, and the physics all agree on.

Every result is optimized around three constraints:

• spatial fidelity,
• SKU fidelity,
• and photometric fidelity.

That is the difference between a concept image and a production asset.

1) Geometry first

Odicon separates scene structure from visual identity.

Spatial grounding

• Depth and normal maps are extracted from the input render.
• Occlusion, floor contact, and perspective are made explicit.
• ControlNet uses these maps so objects land where they can exist.

This avoids the classic failure mode: furniture that looks right but does not belong in the room.

2) Identity preserved

The system does not ask the model to invent a product from scratch.

It searches the catalog first, then conditions generation.

1. Encode catalog references into a multimodal vector index.
2. Retrieve the nearest neighbors for the selected SKU.
3. Inject those references as high-confidence conditioning signals.

The result is consistent fabric, shape, and hardware details, even across long sessions and noisy prompts.

3) Seamless integration

Placement is solved in latent space, not as a hard image paste.

• A soft mask defines the insertion zone.
• Gaussian falloff smooths edge transitions.
• Latent blending merges the synthetic object with scene context before decode.

This produces transitions that read as designed, not composited.

4) Lighting as a constraint

Visual realism is treated as an optical constraint.

• Estimated light direction and intensity are propagated into the diffusion process.
• Material response is adjusted per object type.
• Contact shadows and highlights are generated from geometry-aware depth.

The scene keeps its coherence from corners to shadows.

5) Reliable operations

Odicon runs as an asynchronous queue pipeline, split by responsibility:

• Request layer: validation, tenant context, payload assembly.
• Orchestration layer: queueing, retries, and scheduling.
• Compute layer: GPU workers for constrained diffusion jobs.

This gives clean separation and predictable behavior at scale.

• Non-blocking API responses under heavy workloads.
• Deterministic job lifecycle (queued → processing → completed/failed).
• Independent horizontal scaling across worker pools.
• Safe retries without user-facing failures.

Why this matters

Odicon is not a prompt trick. It is a constrained system for reliable scene synthesis.

• geometry first,
• identity second,
• coherence all the way through.

That is why the output is practical: reviewable, repeatable, and ready for decision-making.

---

Engineering Blueprint (Compact)

Request flow

Designer UI
  -> API layer
    -> validation + context enrichment
    -> job payload creation
      -> task queue
        -> compute worker
          -> constrained generation
            -> asset storage + state update
              -> client notification (stream/webhook)

Queue payload example

{
  "requestId": "od-2026-03-30-a2f9",
  "tenantId": "acme-design",
  "roomImageId": "img_room_9812",
  "skuId": "sku_chair_204",
  "placement": {
    "zone": [0.41, 0.58, 0.27, 0.36],
    "softMaskBlurPx": 24,
    "shadowStrength": 0.62
  },
  "conditioning": {
    "spatial": {
      "controlNet": true,
      "depthScale": 0.9,
      "normalScale": 0.8
    },
    "identity": {
      "adapter": "ip-adapter",
      "kReference": 6,
      "referenceSource": "catalog-embedding-index-v3"
    }
  },
  "output": {
    "format": "png",
    "width": 1536,
    "quality": "high",
    "provenance": true
  }
}