Model Architecture

How DecorViz Thinks

A proprietary AI image synthesis pipeline that turns two unstructured photographs into a single photorealistic composite. No 3D assets. No depth sensors. No room scans.

Monocular Depth Estimation Light Field Analysis Generative Compositing Shadow Synthesis Two-Pass Mirror Rendering Text-Conditioned Control
01 / Full Synthesis Pipeline

From Two Photos to One Photorealistic Output

The DecorViz model processes two inputs in a single end-to-end synthesis pipeline. Each stage conditions the next: segmentation feeds compositing, geometry feeds perspective, photometry feeds shadow synthesis.

ROOM PHOTO RGB image uncalibrated PRODUCT IMAGE any background any source PROMPT optional natural language SEGMENTATION product isolation bg removal SCENE GEOMETRY depth estimation vanishing points PHOTOMETRY light field est. color temp. COMPOSITING perspective align depth placement SHADOW SYNTH occlusion light cast OUTPUT photorealistic composite original resolution STAGE 1 STAGE 2–3 STAGE 4 STAGE 5
02 / Stage 1

Robust Product Segmentation

Before compositing begins, the model runs an automatic segmentation stage that isolates the target product from any input image. The pipeline is engineered to handle the full range of real-world inputs without requiring clean or structured source material.

  • 01 Clean backgrounds. Product on white, grey, or transparent PNG. Highest fidelity baseline.
  • 02 Lifestyle photographs. Product embedded in a styled or staged room scene. Model identifies and extracts the target object from surrounding context.
  • 03 E-commerce screenshots. Mixed-content images containing product photography, promotional banners, UI chrome, navigation, and text. The model locates and isolates the product without manual masking.
WHITE BG LIFESTYLE E-COMMERCE SCREENSHOT clean lifestyle screenshot SEGMENTATION automatic object isolation ISOLATED PRODUCT background removed
03 / Stage 2

Monocular Scene Geometry Inference

From a single uncalibrated 2D photograph, the model performs implicit monocular depth estimation and full scene structure analysis. No stereo input, depth sensor, or structured light is required.

  • 01 Vanishing point inference. Perspective convergence lines are detected and the dominant vanishing point field is reconstructed to establish scene orientation.
  • 02 Floor and ground plane localization. The model identifies the primary support surface: indoor flooring, outdoor patio, decking, or grass. This plane anchors all product placement.
  • 03 Per-pixel depth field. A continuous depth map is inferred across the scene, enabling correct depth-sorted placement of inserted objects without collision with existing geometry.
  • 04 Spatial scale estimation. Relative object sizes within the scene are analyzed to establish a coherent scale reference for correctly sizing inserted products.
VP far mid near FLOOR PLANE product placed at depth Z spatial scale estimation
04 / Stage 3

Photometric Environment Analysis

The model performs implicit light field estimation from the scene photograph, recovering the full photometric profile needed to render the inserted product with illumination-consistent fidelity.

Parameter Description
Light direction Dominant source angle and azimuth inferred from shadow orientation and highlight distribution across surfaces.
Intensity Luminance level of the primary light source, determining shadow contrast and specular response on the inserted product.
Color temperature Kelvin range of the dominant light source recovered from scene white balance. Applied to inserted product surface shading.
Angular distribution Soft vs. hard light inferred from shadow edge sharpness. Determines shadow penumbra width in synthesis.
LIGHT SOURCE dominant direction 2700K warm 6500K cool color temperature recovery intensity product
05 / Stage 5

Shadow and Occlusion Synthesis

Shadows are geometrically derived from the inferred scene, not applied as static overlays or pre-baked assets. Each shadow component is synthesized independently and composited to match the photometric environment.

Contact Shadow
Soft darkening at the object-to-floor boundary. Radius and intensity derived from inferred light distance.
Ambient Occlusion
Occlusion accumulation in concave regions, corners, and recesses adjacent to the placed object.
Directional Cast
Long-form shadow cast across the floor plane, angled and scaled to match the recovered light source direction.
Penumbra
Shadow edge softness scaled by angular distribution of the light source: hard for direct sun, soft for diffuse ambient.
L directional cast contact shadow ambient occ.
06 / Reflective Surfaces

Two-Pass Mirror Rendering

When the scene contains mirrors or highly reflective surfaces, the model applies a two-pass rendering approach to generate physically consistent reflections that incorporate the newly placed product.

  • P1 Pass 1: Scene establishment. Room geometry, lighting, and product placement are finalized. The primary composite is rendered without reflective content.
  • P2 Pass 2: Reflective content generation. Mirror surfaces receive reflected content derived from Pass 1 state: the scene geometry, placed product, and lighting are all visible in the reflection with correct perspective inversion and attenuation.
PASS 1 mirror geometry + lighting established PASS 2 reflected reflection generated from pass 1 state product appears correctly in reflection with perspective inversion and luminance attenuation no post-process overlay -- reflection is synthesized from scene state
07 / Natural Language Interface

Text-Conditioned Scene Control

An optional natural language prompt allows users to direct specific parameters of the generation. Without a prompt, the model applies fully automatic placement, depth positioning, scale inference, and lighting synthesis. When a prompt is provided, it conditions three independent subsystems:

Spatial Placement
The model performs semantic scene understanding, identifying existing objects and their positions, then resolves spatial relationship instructions.
"put the sofa between the TV and coffee table"
"hang the lamp over the dining table"
Scale Calibration
Natural language dimensional descriptions are parsed and applied to constrain the scale calibration system, improving product-to-room proportionality.
"my living room is 4 meters wide"
"ceiling height is around 2.8m"
Scene Attribute Control
Lighting attributes linked to specific scene elements are parsed and applied to the photometric rendering of the relevant fixture and its scene contribution.
"add warm pink light from the floor lamp"
"scene at golden hour lighting"
08 / Product Fidelity

Material and Appearance Preservation

Material properties, surface texture, and colorimetry of the product are preserved through the synthesis process. The model infers surface type from the product image and applies material-specific light response accordingly. Product geometry and surface appearance are not hallucinated or reinterpreted.

Surface Type Rendering Behavior
Specular / Lacquered Specular highlights positioned relative to recovered light source. Environment reflections derived from scene context.
Fabric / Textile Texture-level shading with directional pile behavior. Subsurface scattering for soft-furnishing materials such as velvet and boucle.
Metal Anisotropic reflection response. Brushed vs. polished behavior inferred from texture distribution. Environment map contribution applied.
Matte / Painted Lambertian diffuse shading. Shadow absorption calibrated to surface albedo recovered from product image colorimetry.
Glass / Transparent Transmission and refraction applied relative to scene background. Fresnel response at silhouette edges.
Natural materials (wood, stone, rattan) Grain and texture directionality preserved. Diffuse response with low-level specular component derived from surface finish.
09 / Generalization

Scene and Category Coverage

The model generalizes across scene types and product categories without category-specific fine-tuning or 3D scanning. Any product photograph of sufficient clarity is a valid input.

Scene Types
Residential interiors: living rooms, bedrooms, dining, home office
Open-plan and studio layouts
Outdoor entertaining areas: patios, terraces, covered pergolas
Mixed interior-exterior transitions
Product Categories
All seating: sofas, sectionals, chairs, benches, stools
All tables: dining, coffee, side, console, desk
Rugs, flooring, wall art, mirrors, storage
Lighting fixtures, plants, large outdoor installations

See the Model in Action

Upload a room photo and any product image. Results in seconds.

Try DecorViz Free