Nicholai
86accadc28
Version evolution (SL 1.0→2.0→3.0), team background, no patents, NVIDIA DiffusionRenderer as open-source competitor, dataset landscape (POLAR, SynthLight, etc.), botocore/AWS SDK in privacy app, MetaHuman EULA fix, user data controversy, and DiffusionRenderer ComfyUI integration across all docs.
149 lines
6.6 KiB
Markdown
149 lines
6.6 KiB
Markdown
# Beeble Studio: Technical Analysis
|
|
|
|
An independent technical analysis of the Beeble Studio desktop
|
|
application, examining which AI models power its Video-to-VFX
|
|
pipeline.
|
|
|
|
|
|
## Findings
|
|
|
|
Beeble's product page states that PBR, alpha, and depth map
|
|
generation are "Powered by SwitchLight 3.0." Analysis of the
|
|
application reveals a more nuanced picture:
|
|
|
|
| Pipeline stage | What Beeble says | What the application contains |
|
|
|-------------|-----------------|-------------------------------|
|
|
| Alpha matte (background removal) | "Powered by SwitchLight 3.0" | `transparent-background` / InSPyReNet (MIT) |
|
|
| Depth map | "Powered by SwitchLight 3.0" | Depth Anything V2 via Kornia (Apache 2.0) |
|
|
| Person detection | Not mentioned | RT-DETR + PP-HGNet via Kornia (Apache 2.0) |
|
|
| Face detection | Not mentioned | Kornia face detection (Apache 2.0) |
|
|
| Multi-object tracking | Not mentioned | BoxMOT via Kornia (MIT) |
|
|
| Edge detection | Not mentioned | DexiNed via Kornia (Apache 2.0) |
|
|
| Feature extraction | Not mentioned | DINOv2 via timm (Apache 2.0) |
|
|
| Segmentation | Not mentioned | segmentation_models_pytorch (MIT) |
|
|
| Super resolution | Not mentioned | RRDB-Net via Kornia (Apache 2.0) |
|
|
| PBR decomposition (normal, base color, roughness, specular, metallic) | SwitchLight 3.0 | Architecture built on segmentation_models_pytorch + timm backbones (PP-HGNet, ResNet); proprietary trained weights |
|
|
| Relighting | SwitchLight 3.0 | Proprietary (not fully characterized) |
|
|
|
|
The preprocessing pipeline--alpha mattes, depth maps, person
|
|
detection, face detection, multi-object tracking, edge detection,
|
|
segmentation, and super resolution--is built entirely from
|
|
open-source models used off the shelf.
|
|
|
|
The PBR decomposition model, marketed as part of SwitchLight 3.0,
|
|
appears to be architecturally built from the same open-source
|
|
encoder-decoder frameworks and pretrained backbones available to
|
|
anyone. No physics-based rendering code (Cook-Torrance, BRDF,
|
|
spherical harmonics) was found in the application binary, despite the
|
|
CVPR 2024 paper describing such an architecture. The proprietary
|
|
element appears to be the trained weights, not the model architecture.
|
|
|
|
The name "SwitchLight" does not appear anywhere in the application
|
|
binary, the setup binary, or the Electron app source code. It is a
|
|
marketing name with no corresponding software component.
|
|
|
|
Beeble does acknowledge the use of open-source models in their
|
|
[FAQ](https://docs.beeble.ai/help/faq): "When open-source models
|
|
are included, we choose them carefully." However, the product page
|
|
attributes all outputs to SwitchLight without distinguishing which
|
|
passes come from open-source components.
|
|
|
|
|
|
## Why this matters
|
|
|
|
Most Beeble Studio users use the application for PBR
|
|
extractions--alpha mattes, diffuse/albedo, normals and depth
|
|
maps--not for relighting within the software. The alpha and depth
|
|
extractions use open-source models directly and can be replicated
|
|
for free. The PBR extractions use standard open-source architectures
|
|
with custom-trained weights. Open-source alternatives for PBR
|
|
decomposition (CHORD, RGB-X) now exist and are narrowing the gap.
|
|
See the [ComfyUI guide](docs/COMFYUI_GUIDE.md) for details.
|
|
|
|
The application bundles approximately 48 Python packages, of which
|
|
only 6 include license files. All identified open-source components
|
|
require attribution under their licenses (MIT and Apache 2.0). No
|
|
attribution was found for core components. See the
|
|
[license analysis](docs/LICENSE_ANALYSIS.md).
|
|
|
|
|
|
## Documentation
|
|
|
|
- **[Full report](docs/REPORT.md)** -- Detailed findings with
|
|
evidence for each identified component and architecture analysis
|
|
- **[License analysis](docs/LICENSE_ANALYSIS.md)** -- License
|
|
requirements and compliance assessment
|
|
- **[Methodology](docs/METHODOLOGY.md)** -- How the analysis was
|
|
performed and what was not done
|
|
- **[ComfyUI guide](docs/COMFYUI_GUIDE.md)** -- How to replicate
|
|
the pipeline with open-source tools
|
|
- **[Verification guide](docs/VERIFICATION_GUIDE.md)** -- How to
|
|
independently verify these findings
|
|
- **[Marketing claims](evidence/marketing_claims.md)** -- Archived
|
|
quotes from Beeble's public pages
|
|
|
|
|
|
## Methodology
|
|
|
|
The analysis combined several non-invasive techniques: string
|
|
extraction from process memory, TensorRT plugin identification,
|
|
PyInstaller module listing, Electron app source inspection, library
|
|
directory inventory, and manifest analysis. No code was decompiled,
|
|
no encryption was broken, and no proprietary logic was examined. The
|
|
full methodology is documented [here](docs/METHODOLOGY.md).
|
|
|
|
|
|
## What this is
|
|
|
|
This is a factual technical analysis. The evidence is presented so
|
|
that VFX professionals can make informed decisions about the tools
|
|
they use. All claims are verifiable using the methods described in
|
|
the [verification guide](docs/VERIFICATION_GUIDE.md).
|
|
|
|
|
|
## What this is not
|
|
|
|
This is not an accusation of wrongdoing. Using open-source software
|
|
in commercial products is normal, legal, and encouraged by the
|
|
open-source community. Using open-source architectures with custom
|
|
training data is how most ML products are built. The concerns
|
|
documented here are narrower:
|
|
|
|
1. Marketing language that attributes open-source outputs to
|
|
proprietary technology
|
|
2. Investor-facing claims about a "foundational model" that appears
|
|
to be a pipeline of open-source components with proprietary weights
|
|
3. A CVPR paper describing a physics-based architecture that does
|
|
not appear to match the deployed application
|
|
4. Missing license attribution required by MIT and Apache 2.0
|
|
|
|
5. The CVPR 2024 paper describes SwitchLight 1.0; the shipped product
|
|
is SwitchLight 3.0, which went through at least two "complete
|
|
architecture rebuilds." The physics-based architecture (Cook-
|
|
Torrance, Normal Net, Specular Net) described in the paper may not
|
|
reflect the deployed product.
|
|
|
|
The first and fourth are correctable. The second is a question for
|
|
investors. The third is a question for the research community.
|
|
|
|
|
|
### Competitive landscape
|
|
|
|
The competitive moat described in Beeble's investor materials is
|
|
eroding rapidly. NVIDIA's DiffusionRenderer (CVPR 2025 Oral, open
|
|
source) performs video-to-PBR decomposition and relighting using video
|
|
diffusion models. Multiple research groups have demonstrated that
|
|
synthetic training data (Blender renders of 3D characters with known
|
|
PBR properties) produces results comparable to lightstage-trained
|
|
methods, without requiring proprietary lightstage captures.
|
|
|
|
No patent applications were found for Beeble or its founders related
|
|
to SwitchLight, relighting, or inverse rendering. The CVPR 2024 paper
|
|
has no associated code release.
|
|
|
|
|
|
## License
|
|
|
|
This repository is licensed under
|
|
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
|