From Flat to Deep: Creating Three-Dimensional Feel

The Gap Between 2D and 3D

Luxury watch in alpine sunrise, city skyline at dusk, Mediterranean villa at golden hour—traditional production required travel, permits, weather luck. Budget: prohibitive. Could 2023 AI video generation solve this?

Upload product photo, generate backgrounds—results were technically impressive: mountain scene had atmospheric perspective, city skyline showed depth blur, villa had architectural coherence. But something was wrong.

"Everything looked like it was shot on green screen. The watch never felt in the environment. It looked like a cardboard cutout floating in front of a beautiful painting."

The problems were subtle but fatal:

Contact shadows: The watch cast no grounding shadow on the table, or shadow direction didn't match environmental lighting.

Reflections: The sapphire crystal should have shown sky and mountain reflections, but reflected generic light patterns that didn't correspond to the scene.

Atmospheric interaction: No dust motes in light beams, no depth haze affecting distant objects more. The watch existed in a different dimensional plane than its environment.

Scale consistency: Background elements (trees, buildings) had inconsistent relative sizes. The sense of "how far away is that mountain?" was broken.

200+ generation attempts later, the fundamental limitation remained clear: these models understood 2D composition, not 3D space. They generated beautiful images that failed at the basic task of placing objects in coherent environments.

Project went to traditional production: $67,000 budget, 6-week timeline. The AI "solution" consumed 40 hours and produced nothing usable. The watch never believed it was in the mountains, and neither did the audience.

The Evolution Timeline: From Layered Images to Spatial Understanding

2019: 2D Compositing—Cutouts and Overlays

Early AI "scene composition" was essentially automated Photoshop work. GANs could generate backgrounds and foregrounds separately, but combining them required:

• Manual masking and edge refinement
• Hand-painted contact shadows
• Careful color matching between layers
• Fixed camera angles (no parallax possible)

A "3D feel" required human artists adding depth cues through manual painting. The AI generated components; humans provided spatial coherence.

2021: Parallax Approximation—Fake Depth

Some 2021 systems attempted depth through:

• Separating foreground/midground/background into distinct generation passes
• Applying different motion blur based on "depth"
• Adding atmospheric perspective through post-processing overlays

The results worked for specific scenarios—slow pans across landscapes with clear depth separation. But any complex spatial relationship (objects occluding each other, characters moving through 3D space, camera movement with parallax) revealed the illusion.

Generation times were 10-15 minutes for 5-second clips, making iteration impractical. Creators accepted "flat but beautiful" rather than pursuing true dimensional coherence.

2023: Implicit Depth—Statistical Patterns

Runway Gen-2 and contemporaries showed improvements in implicit depth understanding:

• Better relative scaling of objects
• More consistent atmospheric perspective
• Improved shadow direction (though still often wrong)
• Occasional correct handling of occlusion

But the depth was statistical, not structural. The models learned that "mountains usually go behind trees" and "close objects are bigger than far objects"—but didn't understand why. When scenes deviated from training distribution, depth coherence collapsed.

Complex 3D scenarios remained problematic:

• Moving cameras through cluttered spaces
• Characters interacting with 3D environments (opening doors, sitting on furniture)
• Reflective surfaces showing accurate environment mapping
• Transparent materials with correct refraction

The workaround: avoid these shots. AI video developed a distinctive "look"—shallow depth of field, limited camera movement, simple backgrounds—that compensated for spatial understanding limitations.

2025: Implicit 3D Representation—Structural Understanding

Seedance 2.0's architecture includes implicit 3D scene representation. The Dual-branch Diffusion Transformer doesn't just predict 2D pixels—it maintains understanding of:

Spatial relationships: Objects occupy specific 3D positions relative to each other and the camera.

Physical light transport: Shadows, reflections, and refractions are computed based on 3D geometry, not painted as 2D effects.

Camera motion parallax: Moving the camera produces correct relative motion between near and far objects.

Surface properties: Materials respond to their environment based on physical properties (roughness, metallic, transparency).

This isn't real-time 3D rendering—it's learned 3D understanding encoded in the model's weights. But the results behave correctly in ways that transform creative possibilities.

Seedance 2.0 Solution: Architecture of Depth

How Implicit 3D Works

Traditional diffusion models generate pixels directly from noise, guided by text embeddings. There's no intermediate representation of "what's in the scene"—just a statistical dance toward probable images.

Seedance 2.0's architecture inserts an implicit 3D layer:

1. Input processing: Images, text, and video references are analyzed to extract 3D scene descriptors (rough geometry, light positions, material properties)

2. Scene representation: The Dual-branch Transformer maintains a latent 3D representation alongside the 2D pixel prediction

3. Physical simulation: Light transport, camera projection, and object relationships are computed in this 3D space

4. Pixel generation: The 2D output is rendered from the 3D representation, ensuring physical consistency

The result isn't perfect 3D reconstruction—it's approximate, learned 3D that captures essential spatial relationships for video generation.

Practical Demonstration: Product in Environment

The Challenge: Place a luxury watch on a wooden table in a mountain cabin environment, with natural lighting through windows.

Seedance 2.0 Approach:

Upload reference images:

• Watch product shots (multiple angles for 3D understanding)
• Wooden table texture reference
• Mountain cabin interior reference showing desired lighting

Enable Director Mode and structure the prompt:

SCENE: Mountain cabin interior, afternoon light through windows
SUBJECT: Luxury watch on wooden table, hero framing

SPATIAL_SETUP:
  - Camera: 45° angle, 50mm equivalent, table height
  - Watch: Center frame, 1 meter from window
  - Window: Camera left, casting natural light
  - Background: Cabin interior with depth

DEPTH_CUES:
  - Foreground: Table surface texture, contact shadow
  - Midground: Watch with environmental reflections
  - Background: Soft window view, atmospheric depth

PHYSICAL_PROPERTIES:
  - Watch crystal: Reflects window and interior
  - Metal surfaces: Respond to light direction
  - Wood grain: Catches light across surface
  - Window glass: Slight refraction of exterior view

What Seedance 2.0 generates:

The output shows correct spatial relationships:

• Contact integration: The watch casts a soft shadow on the wood grain, oriented correctly for window light. The wood texture shows appropriate foreshortening.

• Environmental reflections: The watch crystal shows a distorted but recognizable reflection of the window and cabin interior—not generic highlights, but specific environmental features.

• Depth layering: Background elements outside the window show atmospheric haze. Interior elements (chairs, fireplace) scale correctly with distance.

• Camera motion stability: If extended with camera movement, parallax behaves correctly—near objects (watch, table) move more than far objects (window view).

Side-by-Side Comparison: Depth Evolution

Native 2K: Where Depth Detail Lives

Depth perception relies on fine detail:

• Texture gradients: Wood grain, fabric weave, stone surfaces that compress with distance
• Edge definition: Sharp near edges, soft far edges
• Micro-shadows: Small surface details casting tiny shadows that create 3D texture
• Specular highlights: Reflections that shift with surface curvature

At 720p, these cues are compressed into ambiguity. Native 2K preserves the gradients that communicate depth:

• Individual wood grain lines show foreshortening
• Fabric texture maintains detail at distance
• Surface imperfections create micro-shadows
• Curved surfaces show highlight gradients

The difference between "flat" and "deep" often comes down to whether these fine cues are preserved or lost.

Director Mode: Controlling 3D Space

The Internal Shot List enables explicit 3D control:

SHOT_1:
  Camera_position: [x: 0, y: 1.2, z: 2.0]
  Look_at: [x: 0, y: 0.8, z: 0]
  Focal_length: 50mm

  Subject_position: [x: 0, y: 0.8, z: 0]
  Subject_rotation: [y: 15°]

  Environment:
    Type: Mountain cabin
    Light_source: Window_left
    Atmosphere: Dust_motes_visible

SPATIAL_CONSTRAINTS:
  - Maintain subject scale across camera movement
  - Preserve contact shadows with surface
  - Environmental reflections must match scene
  - Background depth_haze proportional to distance

Seedance 2.0 interprets these constraints through its implicit 3D representation, generating output that respects spatial relationships.

Speed Enables Depth Exploration

Creating depth-coherent scenes traditionally required trial and error. With 29-second generation times, you can:

1. Generate with basic depth setup
2. Review for spatial coherence issues
3. Adjust camera angle or subject position
4. Regenerate and compare
5. Iterate until depth "feels right"

This process might take 10-15 minutes with Seedance 2.0. With 4-5 minute generation times, it would take 1-2 hours—and you'd settle for "good enough" instead of "actually coherent."

You Can Act Now: Building Spatially Coherent Scenes

Step 1: Provide 3D Information Through References

Seedance 2.0 extracts spatial understanding from:

• Multiple angles of the same object: Upload 3-4 views of your subject to establish 3D form
• Environment references: Images showing desired depth relationships
• Lighting references: Photos demonstrating how light interacts with space

The more 3D information you provide, the better the spatial coherence.

Step 2: Use This Depth-Focused Prompt Template

SPATIAL_CONCEPT: [Overall 3D arrangement]

CAMERA:
  Position: [Relative to scene]
  Height: [Eye level/looking up/looking down]
  Movement: [Static/pan/dolly/etc]

SUBJECT_PLACEMENT:
  Position: [In 3D space]
  Orientation: [Facing direction]
  Contact: [How subject touches environment]

DEPTH_LAYERS:
  Foreground: [Close elements with detail]
  Midground: [Primary subject and immediate environment]
  Background: [Distant elements with atmosphere]

LIGHTING_DEPTH:
  Source: [Where light comes from]
  Quality: [How it wraps around forms]
  Shadows: [Direction and softness]

REFLECTIONS/REFRACTIONS:
  - [How surfaces interact with environment]

CONSISTENCY_CHECKS:
  - Scale relationships
  - Shadow directions
  - Contact integration
  - Parallax behavior

Step 3: Review for Depth Coherence

Before accepting generated output, check:

• Contact points: Does the subject cast appropriate shadows on surfaces?
• Reflections: Do reflective surfaces show environment-appropriate imagery?
• Scale: Do distant objects look appropriately smaller than near ones?
• Atmosphere: Is there depth-appropriate haze or clarity?
• Motion: If camera moves, does parallax behave correctly?

If any check fails, adjust and regenerate. Speed makes this iteration practical.

12-Month Prediction: The Depth Horizon

Q2 2026: Explicit depth map input. Provide rough depth paintings or 3D proxies; Seedance 2.0 generates video respecting that geometry.

Q3 2026: Volumetric effects control. Specify fog density, light beam scattering, atmospheric particles with spatial precision.

Q4 2026: Reflection probe emulation. Upload environment HDRIs or 360° captures; reflective surfaces respond accurately to that specific environment.

2027: Hybrid workflows. Combine AI-generated elements with real-time 3D renders, maintaining coherent lighting and depth between both.

---

Series Navigation

Previous: E08: From Slow to Fast Next: E10: From Static to Motion

---

Depth isn't just a technical achievement—it's the foundation of presence. When objects believe they're in space, the audience believes they're witnessing reality. What worlds will you build when your canvas has three dimensions?