StyleAlign: Analysis and Applications of Aligned StyleGAN Models
Authors: Zongze Wu, Yotam Nitzan, Eli Shechtman, Dani Lischinski
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | First, we empirically analyze aligned models and provide answers to important questions regarding their nature. In particular, we find that the child model s latent spaces are semantically aligned with those of the parent, inheriting incredibly rich semantics, even for distant data domains such as human faces and churches. Second, equipped with this better understanding, we leverage aligned models to solve a diverse set of tasks. In addition to image translation, we demonstrate fully automatic cross-domain image morphing. We further show that zero-shot vision tasks may be performed in the child domain, while relying exclusively on supervision in the parent domain. We demonstrate qualitatively and quantitatively that our approach yields state-of-the-art results, while requiring only simple fine-tuning and inversion. |
| Researcher Affiliation | Collaboration | Zongze Wu The Hebrew University Yotam Nitzan Tel-Aviv University Eli Shechtman Adobe Research Dani Lischinski The Hebrew University |
| Pseudocode | No | The paper does not contain any pseudocode or explicitly labeled algorithm blocks. |
| Open Source Code | Yes | Separately, source code and pretrained models have been made available in the project s repository. |
| Open Datasets | Yes | We transfer a parent Style GAN2 model pretrained on FFHQ to the Mega cartoon dataset (Pinkney & Adler, 2020) and to AFHQ dog faces dataset (Choi et al., 2020), using ADA (Karras et al., 2020a). |
| Dataset Splits | No | The paper mentions using well-known datasets like FFHQ, Mega, and AFHQ for fine-tuning and evaluation. However, it does not explicitly provide the training/validation/test splits used for its specific experiments or evaluations. |
| Hardware Specification | Yes | creating a low-resolution model (512 512) in this way is computationally efficient, requiring less than 2 days of fine-tuning on a single GTX1080Ti GPU |
| Software Dependencies | No | The paper mentions 'tensorflow', 'Style GAN2', and 'Style GAN2-ADA implementations' but does not specify exact version numbers for these software dependencies. |
| Experiment Setup | Yes | Given a model pretrained on the parent domain, we fine-tune it on the child domain. Specifically, we use model config-f and the default hyper-parameters from the official Nvidia Style GAN2 and Style GAN2-ADA implementations in tensorflow. We use the augmentations of Style GAN2-ADA only when the child domain is AFHQ or Metface. |