Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Low-Rank Head Avatar Personalization with Registers
Authors: Sai Tanmay Reddy Chakkera, Aggelina Chatziagapi, Md Moniruzzaman, Chen-Ping Yu, Yi-Hsuan Tsai, Dimitris Samaras
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To demonstrate the efficacy of our personalization method, we collect a dataset of talking videos of individuals with distinctive facial details, such as wrinkles and tattoos. Our approach faithfully captures unseen faces, outperforming existing methods quantitatively and qualitatively. Our method outperforms state-of-the-art approaches, like meta-learning and vanilla Lo RA, both quantitatively and qualitatively, while it only requires a small number of parameters to adapt. |
| Researcher Affiliation | Collaboration | Sai Tanmay Reddy Chakkera Department of Computer Science Stony Brook University EMAIL Aggelina Chatziagapi Department of Computer Science Stony Brook University EMAIL Md Moniruzzaman Atmanity Inc. EMAIL Chen-Ping Yu Atmanity Inc. EMAIL Yi-Hsuan Tsai Atmanity Inc. EMAIL Dimitris Samaras Department of Computer Science Stony Brook University EMAIL |
| Pseudocode | No | The paper describes its methodology in Section 3, 'Proposed Method', using descriptive text and mathematical equations, but it does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: We plan to release the data and code to the public upon acceptance. Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [No] Justification: We will release our model and data upon acceptance with appropriate licence. |
| Open Datasets | Yes | In addition to Rare Face-50, we use VFHQ test set and HDTF dataset to evaluate our method. HDTF (Zhang et al., 2021) consists of 362 videos each cropped and resized to 512x512 resolution. VFHQ Test consists of 50 high quality videos from 50 different identities cropped and resized to 512 512 resolution. Each video is around 4 to 10 seconds in diverse poses and settings. |
| Dataset Splits | Yes | VFHQ Test consists of 50 high quality videos from 50 different identities cropped and resized to 512 512 resolution. Each video is around 4 to 10 seconds in diverse poses and settings. In Table 6, we ablate on the length of the videos used to adapt our head avatars. We see that reducing the length of the adaptation video causes a drop in the performance of the method. Note that we trim the original videos to the first 4, 2, and 1 second for these experiments. |
| Hardware Specification | Yes | Our adaptation takes 35 minutes on an RTX A5000 GPU, consuming 23GB of VRAM. This adaptation takes 25 minutes on an RTX A5000 GPU, consuming 14.9GB of VRAM. Given resource constraints, we implement a single GPU version of Reptile (Nichol and Schulman, 2018), thus taking 12 days to complete the pretraining task on a Quadro RTX 8000 GPU, consuming 45GB of VRAM. After the pre-training task, we adapt the model on an identity for 120 steps with the same learning rate as the inner loop, which takes 4 minutes on an RTX A5000 GPU consuming 14.9GB of VRAM. |
| Software Dependencies | No | We use min Lo RA (Chang and Kelly) library to add Lo RA (Hu et al., 2022) parameters to all layers of a pretrained pytorch model. We use Adam (Kingma and Ba, 2017) optimizer with learning rate set to 1e 4 for the Lo RA layers whereas the learning rate is set to 1e 3 for parameters in the Register Module. |
| Experiment Setup | Yes | We initialize embeddings e using Xavier Normal initialization (Glorot and Bengio, 2010). We adapt head avatars with our method for a total of 1000 iterations. The batch size is set to 2. We use Adam (Kingma and Ba, 2017) optimizer with learning rate set to 1e 4 for the Lo RA layers whereas the learning rate is set to 1e 3 for parameters in the Register Module. We use a linear learning rate scheduler with a start factor of 1.0 and an end factor of 0.1 at the 1000th iteration. |