Martin Nicolas Everaert
Research Subjects
Martin Nicolas Everaert has a total of 5 co-authors. Here are the top co-authors by number of shared publications:
- Marco Bocchio
Largo.ai,Lausanne,Switzerland
2 shared publications - Radhakrishna Achanta
EPFL,School of Computer and Communication Sciences,Switzerland
2 shared publications - Sabine Süsstrunk
EPFL, Lausanne, Switzerland
2 shared publications - Sami Arpa
Largo.ai,Lausanne,Switzerland
2 shared publications Athanasios Fitsios
EPFL,School of Computer and Communication Sciences,Switzerland
1 shared publication
Latest Publications
- Open Access
- 3 January 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 4013-4022
Exploiting the Signal-Leak Bias in Diffusion Models
There is a bias in the inference pipeline of most diffusion models. This bias arises from a signal leak whose distribution deviates from the noise distribution, creating a discrepancy between training and inference processes. We demonstrate that this signal-leak bias is particularly significant when models are tuned to a specific style, causing sub-optimal style matching. Recent research tries to avoid the signal leakage during training. We instead show how we can exploit this signal-leak bias in existing diffusion models to allow more control over the generated images. This enables us to generate images with more varied brightness, and images that better match a desired style or color. By modeling the distribution of the signal leak in the spatial frequency and pixel domains, and including a signal leak in the initial latent, we generate images that better match expected results without any additional training.
- 3 January 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 4013-4022
Open AccessExploiting the Signal-Leak Bias in Diffusion Models
There is a bias in the inference pipeline of most diffusion models. This bias arises from a signal leak whose distribution deviates from the noise distribution, creating a discrepancy between training and inference processes. We demonstrate that this signal-leak bias is particularly significant when models are tuned to a specific style, causing sub-optimal style matching. Recent research tries to avoid the signal leakage during training. We instead show how we can exploit this signal-leak bias in existing diffusion models to allow more control over the generated images. This enables us to generate images with more varied brightness, and images that better match a desired style or color. By modeling the distribution of the signal leak in the spatial frequency and pixel domains, and including a signal leak in the initial latent, we generate images that better match expected results without any additional training.
- Open Access
- 1 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 2251-2261
Diffusion in Style
We present Diffusion in Style, a simple method to adapt Stable Diffusion to any desired style, using only a small set of target images. It is based on the key observation that the style of the images generated by Stable Diffusion is tied to the initial latent tensor. Not adapting this initial latent tensor to the style makes fine-tuning slow, expensive, and impractical, especially when only a few target style images are available. In contrast, fine-tuning is much easier if this initial latent tensor is also adapted. Our Diffusion in Style is orders of magnitude more sample-efficient and faster. It also generates more pleasing images than existing approaches, as shown qualitatively and with quantitative comparisons.
- 1 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 2251-2261
Open AccessDiffusion in Style
We present Diffusion in Style, a simple method to adapt Stable Diffusion to any desired style, using only a small set of target images. It is based on the key observation that the style of the images generated by Stable Diffusion is tied to the initial latent tensor. Not adapting this initial latent tensor to the style makes fine-tuning slow, expensive, and impractical, especially when only a few target style images are available. In contrast, fine-tuning is much easier if this initial latent tensor is also adapted. Our Diffusion in Style is orders of magnitude more sample-efficient and faster. It also generates more pleasing images than existing approaches, as shown qualitatively and with quantitative comparisons.