Just released a simple open-source tool that uses classical ML to blend regions from multiple 3D meshes into new ones
Pick source meshes, paint regions, interpolate in latent space.
Great for learning how ML can work with geometry.
👉 github.com/lopezmauro/m...
#ml #3d #opensource #techart
Big thanks to Jeff Brodsky of Vertex Theory for this incredible video showcasing my tool! If you haven’t seen it yet, don’t miss out, it’s a fantastic walkthrough! 🎥 youtu.be/kEQK5Sd4484?...
🚀Repo Update!
Based on your feedback, I’ve added:
✨ Import/Export functionality
⏰ Custom train time range (for non-animated inputs)
🔄 Linear Regression often struggles with Euler rotations. I’ve added rotation matrix support
github.com/lopezmauro/m...
#MachineLearning #AutodeskMaya #ML #Rigging
I created these nodes in Maya since most production rigs|animations are done there. There aren’t many production-level character rigs in Houdini.
That said, the inference math quite is simple, so I might put together a few snippets to achieve something similar in Houdini in the future. Stay tuned😉
👍 Give a like if you'd love a follow-up on adding non-linearity to linear regression! Or drop a comment to let me know what you'd like to see next.
I’m always excited to dive deeper into #MachineLearning applied to #Rigging topics and #CharacterTD
Excited to share my PCA Blendshape Node for #AutodeskMaya
This educational tool lets you explore PCA (Principal Component Analysis), allowing you to choose how much to compress your data while optimizing performance and preserving detail.
github.com/lopezmauro/m...
#MachineLearning #ML #Rigging
💡 Coming next: a PCA Blendshape Node
These tools are instructional—Not meant for production use but great for exploring ML in 3D animation. Feedback & contributions are welcome!
🤖 Dive into ML in Autodesk Maya
I’ve made a repo with example ML nodes for Maya, starting with a Regression Node for linear regression inference. Perfect for learning & experimentation!
📂 Check it out: github.com/lopezmauro/m...
#MachineLearning #AutodeskMaya #ML #Rigging #CharacterTD
You think you know… but you have no idea. Naughty Dog’s 5th franchise is our wildest adventure. We cannot wait to show you more!
Super proud and happy to be able to finally show this project and showcase the in-game Realtime ML facial and cloth tech that I've working on 🤓
#Rigging #MachineLearning #ML #CharacterTD #Intergalactic #NaughtyDog
youtu.be/o7TVPoxwi74?...
Also, in my experience, ML models are easier to implement in custom pipelines. Their inference relies on straightforward linear algebra, making them simpler implement. And since they are easy to interpret, can be integrated with post train fine-tune with artist-driven tools (edit weight and bias)
Neural networks are great for model complex, non-linear relationships. However, for simpler problems, ML's advantages in efficiency (like you mention), data requirements (they can be trained with smaller data sets), and interpretability (easier to understand how each feature affect the outputs).
3️⃣ PCA for Generating New 3D Faces (Eigenfaces for Meshes)
Use PCA to reduce the dimensionality of 3D facial data, then generate new faces by interpolating in the reduced space, and then convert it back to the original space for a brand new face.
2️⃣ K-Means for Vertex Clustering by Position and Normals
Use k-means clustering to group vertices based on spatial proximity and normal direction. This helps segment parts of a mesh automatically (e.g., head, torso, limbs).
1️⃣ Linear Regression for "Smart" Driven Keys:
Use linear regression to model the relationship between a driver (e.g., joint rotation) and a driven (e.g., blendshape weight). Instead of setting keys manually, train the relationship using example data and let the model interpolate for unseen values.
Neural networks might be stealing the spotlight, but classical machine learning is still a powerful and accessible tool for smarter, faster rigging workflows. In this thread, I’ll share three cool ML examples you can use right now 🧵👀
#Rigging #MachineLearning #ML #CharacterTD
Here are some pseudo-code images.
🎯 Modular and easy to extend:
✅ Add constraints with loss functions.
✅ Assign weights to losses for flexibility.
✅ Add new types of losses dynamically.
Perfect for tasks like mesh fitting, alignment, or other approximations
#TechInnovation
#Rigging
#CharacterTD
Neural networks can be powerful tools for approximating mesh deformations by aligning a source mesh to a target mesh
One key advantage is the flexibility to include various constraints via loss functions such as:
NORMAL LOSS: Preserve source normals to maintain the overall shape of the mesh. EDGE LOSS: Preserve edge lengths to retain the original volume of the mesh
SYMMETRY LOSS: If the source mesh has symmetry, can penalizes asymmetrical deformations. LAPLACIAN LOSS: minimizes differences in vertex neighborhoods, preventing unwanted distortions
💡 Did you know? Neural network libraries like PyTorch can act as general-purpose approximators without building a network.
Just optimize a tensor with losses to solve tasks such a mesh matching with constraints. Continue on🧵
#MachineLearning
#NeuralNetworks
#TechInnovation
#Rigging
#CharacterTD
Locked in the zone, working hard... then I feel it on my neck: the undeniable presence behind me. The Attention Seeker has arrived.
Normalize inputs for stable training, but not outputs—non-normalized outputs reduce approximation errors.
Avoid Euler angles due to gimbal lock. Quaternions are better but can be inconsistent. Rotation matrices are the most stable.
predict only 6 values of the rotation matrix and compute the last 3 with a cross-product. This keeps the network smaller
While PReLU works well for large networks, Tanh provides more stable predictions outside the training data.
Key Takeaways for Neural Networks Driving Joints:
After experimenting with neural networks for joint rotations, here are a few lessons I’ve learned
Have you tried similar methods? Let’s discuss!
#MachineLearning #DeepLearning #NeuralNetworks #AnimationTech #Rigging
