https://github.com/robinka/freq-domain-ar
Autoregressive frequency-domain image generation
https://github.com/robinka/freq-domain-ar
ai auto-regressive decoder fourier frequency genai image-generation lightning python spectrum torch transformer
Last synced: 2 months ago
JSON representation
Autoregressive frequency-domain image generation
- Host: GitHub
- URL: https://github.com/robinka/freq-domain-ar
- Owner: RobinKa
- Created: 2025-03-23T12:39:48.000Z (3 months ago)
- Default Branch: master
- Last Pushed: 2025-03-23T19:13:36.000Z (3 months ago)
- Last Synced: 2025-03-23T19:21:43.327Z (3 months ago)
- Topics: ai, auto-regressive, decoder, fourier, frequency, genai, image-generation, lightning, python, spectrum, torch, transformer
- Language: Python
- Homepage:
- Size: 457 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Autoregressive frequency-domain image generation
Uses a decoder-only transformer to autoregressively generate images in the frequency domain.
## How it works
Here are some bulletpoints on how it works:
- A few invertible transformations are done on the image to put them into frequency domain in a nice format for autoregressive prediction. These need to be undone again for plotting.
- fourier-transform without redundancy using rfft2
- log1p is applied to bring the values into a reasonable range
- flatten and sort by the magnitude of the frequency, because we want to start predicting the low frequencies first
- prepend the label (0-9 for mnist) so we know what to generate
- causal decoder-only transformer is trained in the usual way
- the model uses a size 15 patchification layer
The current implementation uses mnist as a dataset. Note that no time was spent optimizing the architecture etc. and this is more of a proof of concept.
## Examples
Below are some examples output where we autoregressively sample 15 pixels at a time (because that's the patchification size). The left part is the frequency domain.
https://github.com/user-attachments/assets/2c3d4b83-f5b8-499e-bcbd-0bc6cb524cf1
https://github.com/user-attachments/assets/697cae57-2faa-46a9-9966-c14fd63a3927
## How to run
- `uv sync` to install dependencies
- `python -m freq_ar.train --log_every_n_steps 1000` to run
Here's how the loss looks during training
