Metoder för Bildgenerering från Latenta Vektorrum: En Jämförande Studie
Abstract
The current boom in generative AI is due to recent advancements in embedding and generation methods as well as increased computational power. For image generation, these techniques first find a small latent space representing the meaningful features of the image. Generation can then proceed with multiple types of neural networks, such as diffusion models or 'generative pre-trained transformers' (GPTs). In this project, several of the state-of-the-art techniques are implemented and compared. The autoencoder latent space is discretized using 'finite scalar quantization' (FSQ), and generation is implemented through a GPT, a 'Gaussian diffusion model', and a 'quantized diffusion model'. The autoencoder model reconstructions achieve an FID of 35.80 on the test dataset, and the generated images reach an FID of 54.89. Autoencoder reconstruction quality decreases as the compression factor of the original image increases. The big separator in terms of image quality of generated images is the method of generation, not the number of model parameters. Both diffusion models are sensitive to the size of the latent space, and performs better on a smaller latent space. Contrarily, the GPT produces the best results on the bigger latent space, from which the autoencoder reconstruction fidelity is better. The GPT is also the generation technique that performs best overall, followed by Gaussian diffusion and lastly, quantized diffusion. The dataset consists of 260K colour images of electrical…
Read the paper
Related studies
- Learning from First-Generation Qualitative Approaches in the IS Discipline: An Evolutionary View and Some Implications for Authors and Evaluators (PART 1/2)
- Commentary on the clinical and preclinical dosage limits of interstitially administered magnetic fluids for therapeutic hyperthermia based on current practice and efficacy models
- Prosthesis: New Technological Opportunities and Innovative Biomedical Devices
- The history and future of neural modeling for cochlear implants
- Ryanodine receptors: waking up from refractoriness
- Novel Transgenic Rabbit Model Sheds Light on the Puzzling Role of Matrix Metalloproteinase-12 in Atherosclerosis
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.