16-825 Assignment 4¶
1. 3D Gaussian Splatting¶
1.1 3D Gaussian Rasterization (35 points)¶

1.2 Training 3D Gaussian Representations (15 points)¶
| Parameter | Learning Rate (lr) |
|---|---|
| opacities | 0.0008 |
| scales | 0.001 |
| colours | 0.025 |
| means | 0.00015 |
| Metric | Value |
|---|---|
| Iterations | 1000 |
| Mean PSNR | 27.843 |
| Mean SSIM | 0.920 |
1.3 Extensions (Choose at least one! More than one is extra credit)¶
1.3.1 Rendering Using Spherical Harmonics (10 Points)¶
Old Output
New Output
📸 Rendering Comparison: New (Q1) vs. Old (Q1.1.5)¶
| New Rendering (Q1/output/q1_render/) | Old Rendering (Q1/output/q1.1.5_render/) |
|---|---|
| Frame 000 | Frame 000 |
![]() |
![]() |
| Frame 018 | Frame 018 |
![]() |
![]() |
| Frame 029 | Frame 029 |
![]() |
![]() |
The most prominent difference is the increased brightness of the newer rendering, which results in the couch shifting to a much lighter green hue when contrasted with the darker tones of the older rendering.¶
2. Diffusion-guided Optimization¶
2.1 SDS Loss + Image Optimization (20 points)¶
2.2 Texture Map Optimization for Mesh (15 points)¶
🌐 3D Mesh Generation Comparison
| Original Mesh (Reference) | Prompt 1: A Hamburger | Prompt 2: Disco Lights | Prompt 3: Multi Coloured Lego |
|---|---|---|---|
|
|
|
|
2.3 NeRF Optimization (15 points)¶
1️⃣ Prompt: A standing corgi dog 🐕
| Example Image | RGB Video | Depth Video |
|---|---|---|
|
2️⃣ Prompt: A red racing car 🏎️
| Example Image | RGB Video | Depth Video |
|---|---|---|
|
3️⃣ Experimental Prompt: A f1 car on track (Dosent Work well)
| Example Image | RGB Video | Depth Video |
|---|---|---|
|
2.4 Extensions (Choose at least one! More than one is extra credit)¶
2.4.3 Variation of implementation of SDS loss (10 points)¶
These results were implemented using L2 Loss¶
1️⃣ Prompt: A standing corgi dog 🐕
| Example Image | RGB Video | Depth Video |
|---|---|---|
|
2️⃣ Prompt: A red racing car 🏎️
| Example Image | RGB Video | Depth Video |
|---|---|---|
|
💡 Loss Model Comparison Summary¶
| Loss Model | Training Details (10,000 Iterations) | Key Observations |
|---|---|---|
| SDS Loss | Trained for 10,000 iterations. | Produced cleaner outputs with significantly less noise. |
| Pixel Space Loss | Trained for 10,000 iterations. | Slightly slower per iteration and appears to require more training iterations to achieve results comparable to the SDS model's output quality. |
Conclusion¶
The SDS Loss Model provided superior visual quality within the same training budget (10,000 iterations), yielding outputs that were much cleaner and less noisy than the Pixel Space Loss Model. The Pixel Space Loss Model is likely underexposed and needs further training to reach comparable results.













