DreamEdit3D: Personalization of Multi-View Diffusion Models for 3D Editing (2025)

This project presents DreamEdit3D, a novel framework for personalized 3D scene editing by leveraging multi-view diffusion models. The approach enables users to edit 3D scenes with fine-grained control through personalized text-driven modifications while maintaining multi-view consistency.

Technologies: Multi-View Diffusion, 3D Gaussian Splatting, Personalization, Deep Learning, PyTorch

• Multi-View Consistency: Ensures coherent edits across all viewpoints by personalizing multi-view diffusion models, avoiding the inconsistencies common in single-view editing approaches.
• Personalized Editing: Enables subject-driven 3D editing by fine-tuning diffusion models on user-provided reference images, allowing precise insertion and modification of objects in 3D scenes.
• 3D Reconstruction Integration: Combines edited multi-view outputs with 3D Gaussian Splatting for high-quality, real-time renderable 3D scene reconstruction.

Paper under submission to ECCV 2026 — GIFs and demo video coming soon.

Project Page →

GenSMPL: Generative Skinned Multi-Person Linear Model (2025)

GenSMPL is a data-driven framework for updating and extending the SMPL body model to better represent diverse body shapes and structural variations of children not captured by the original model. Our approach focuses on modifying key components such as the identity PCA space to adapt to entirely new categories.

Technologies: SMPL, PCA, 3D Body Modeling, Dense Registration, Synthetic Data Generation, PyTorch

• Controlled 3D Data Generation: Generating controlled 3D children's body data to serve as training input for learning new shape priors.
• Dense Registration: Performing dense registration to the SMPL template, ensuring consistent mesh topology across all generated samples.
• Shape Prior Learning: Learning new shape priors from aligned meshes via PCA, extending the generalization capability of SMPL to children categories.

Validated on synthetic and real-world datasets, achieving improved reconstruction accuracy and more realistic shape representation.

Paper under submission — visuals and demo coming soon.

Demo: Try GenSMPL on Hugging Face

Geometric Scene Understanding (2024)

This project investigates the generalization potential of state-of-the-art computer vision models trained using the Gaussian Splatter Method. Two primary scenarios were explored:

Technologies: Gaussian Splatter, Diffusion, Deep Learning

• Joint Training: The model was trained jointly on datasets including NMR_Cars, Sofa, and Bench. This training strategy resulted in strong generalization to NMR_Chairs, demonstrating the model's robustness and adaptability across similar but distinct object categories.
• Fine-tuning Pre-trained Models: A model pre-trained on SRN_Cars was fine-tuned on Sofa and Bench datasets. This approach also yielded good generalization on NMR_Chairs, showcasing the effectiveness of fine-tuning strategies in enhancing model performance on new categories.
• Dataset Source: ShapeNet

Visualizations:

The tables below compare ground truth (GT) images with the corresponding rotated (ROT) images before and after the application of the generalization techniques.

GT (Joint Training)
ROT (Joint Training)

GT (Before Fine-tuning)
ROT (Before Fine-tuning)

GT (After Fine-tuning)
ROT (After Fine-tuning)

Conclusion:

• Both models struggle when trained on joint categories that are not similar, evident from lower PSNR values.
• When categories are similar, accuracy improves significantly with higher PSNR values.
• Using datasets with low relevance decreases model performance.

3-Dimensional Reconstruction of a Room based on Stereo Images (2023)

Technologies: MATLAB, Multiview Geometric Reconstruction

MMOM System Fullstack Development (2023 to 2024)

• This project represents the culmination of 15 months of dedicated work during my tenure on the TOA team.
• Due to confidentiality clauses and agreements with TUM, I am unable to provide specific details or share any gifs, videos, or pictures.