nev
@neverrixx
Followers
623
Following
948
Media
78
Statuses
290
Joined March 2019
We quantized all Gemma Scopes into 4 bits, reducing memory and storage requirements by about 4 times with ~20% higher variance unexplained. You can try the quantized versions and make your own quantized SAEs with DM me if you see bugs or have requests.
Sparse Autoencoders act like a microscope for AI internals. They're a powerful tool for interpretability, but training costs limit research. Announcing Gemma Scope: An open suite of SAEs on every layer & sublayer of Gemma 2 2B & 9B! We hope to enable even more ambitious work
1
3
28
In addition, there are some other ways to deal with this:.1) use less dimensions.2) increase the maximum noise level.3) use more diffusion steps during inference.(9/8).
0
0
8
Cross Labs fine-tuned Stable Diffusion using this technique. The result is much better at handling color variation across scenes. You can read more about the results and find the model here (7/7).
1
2
24
This gives us the understanding necessary to solve this problem. The author chose something simple: adding the lowest frequency to the noise. This makes it correlated, and the model learns to notice the change and modify the average color more quickly. (6/7)
1
0
8
The average color is determined by the lowest frequency - the one with an infinite wavelength. Diffusion with a finite noise level can never meaningfully change the average color! The average of the image is the same as that of the starting noise. (5/7)
1
2
10
The real reason this happens has to do with frequencies. When noise is added to an image, the higher frequencies are lost first, and low frequencies - the general structure of the image - change slowly. (4/7)
1
0
4
Let's train a few toy diffusion models on Gaussian noise offset by some constant - here it's 2. If the data is low-dimensional, the task is easy. But the more dimensions the data has, the harder it is to fit the real mean. (3/7)
1
0
3
Most images generated by Stable Diffusion have average pixel values close to 0.5. This is not a problem specific to the data - even you try fine-tune it on a pure black image for 3000 steps, the model completely fails. (2/7)
1
1
7
Why does Stable Diffusion struggle with making very dark or light images? This problem actually affects all diffusion models - it's impossible to fit a purely black or white image!.A simple fix exists. This is a thread about a blog post from Cross Labs. (1/7)
7
26
189
