Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
An extremely good ASCII art generator, based on machine learning
| Date | Stars |
|---|---|
| 2026-07-31 | 259 |
| 2026-08-03 | 259 |
| 2026-08-06 | 261 |
Today
+2 stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# Subpixel ASCII Art via Gradient Descent An approach to ASCII art generation that uses gradient descent optimization instead of lookup tables. This project produces high-quality ASCII art by treating character selection as a differentiable optimization problem. Is it overkill? Yes but fuck it we ball ⚠️⚠️⚠️⚠️ **this readme is vibed by claude so don't judge me if it's written in a npc like way** ⚠️⚠️⚠️⚠️ ## Gallery ### [Bad Apple](https://x.com/gf_256/status/2015012967427854573) EXPERIENCE IT LIVE: ``` telnet bad.apple.zellic.io -= or =- nc bad.apple.zellic.io 23 -= or =- ssh [email protected] ``` or https://bad.apple.zellic.io YES IT SUPPORTS ANIMATIONS WITH TEMPORAL REGULARIZATION LOSS ---    --- Demonstration of psychovisual tuning  - Diversity loss weight: penalty for using the same characters over and over again (adds noise = looks better. less banding etc) - Multiscale loss weight: also downscale the learned and target image and compute loss over that too (makes it do dithering) --- Default settings <img src="testcases/test2.png" width="378">  --- Diversity weight 0.0 and disable annealing: <img src="testcases/test.png" width="378"> <img src="examples/zellic.png" width="378"> Diversity weight 0.02 and 0.01:   --- Benchmarks / torture tests <img src="testcases/test3.png" width="378">  <img src="testcases/test4.png" width="378">  --- Identity mapping test  ## Warping Learnable affine transformation (scale+translate), warp matrix. Is this totally overkill? Yes! Does it make results significantly better? Yes! Without warping enabled (only scaling and translation), we struggle to learn an identity mapping because the input raster are misaligned with the character grid.  With warping enabled, we successfully learn significant part of the identity mapping. This is not easy!  Warping significantly improves reconstruction loss by leveraging "row gaps" as part of the pattern allowing for cleaner horizontal and vertical edges. In practice even substantial warping is not noticeable in the final result.     Pay attention to the lips, bottom edges of the eyes, bottom of the chin, and eyebrows. They become aligned with the character grid so the row gaps can be used to create clean horizontal edges. Pay attention to the split in the middle of the bangs, the right side of the face near the temple, and various vertical lines in the hair. This becomes aligned with a column boundary to create a clean vertical edge.   ## Dynamic contrast (and dark mode) Learnable contrast curve per each point on the character grid is bilinearly interpolated. Dark mode is also supported as shown in this image  Yes this is completely overkill and it's mostly an artistic effect to add a fuck ton of shading like those old 90s ansi files. ``` Curve Non-linearity Map (deviation from identity):
Excerpt of 19,655 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:c234a19e6c6185a0, llm:description: 'An extremely good ASCII art generator, based on machine learning'; topics: ascii-art, machine-learning
matched fp:c234a19e6c6185a0, llm:description: 'An extremely good ASCII art generator, based on machine learning'; topics: ascii-art, machine-learning
matched fp:c234a19e6c6185a0, llm:description: 'An extremely good ASCII art generator, based on machine learning'; topics: ascii-art, machine-learning