FreeU support

[lpn256]

FreeU support

[bssrdf]

I gave a try but I am not sure if implemented correctly. Here are the comparison without and with FreeU.

FreeU OFF	FreeU ON

Above ones are all using artiusV21_v21.safetensors file. Below comparison switched to v2-1_768-nonema-pruned.safetensors.

FreeU OFF	FreeU ON

FreeU definitely won here 😸 .

[FSSRepo]

Interesting, I'll see if I can give it some time to at least try if it works.

Bad new:

To implement this, it would be necessary to create the operation for the fourier transform and its inverse in ggml, and that is very difficult.

[bssrdf]

@FSSRepo, @leejet, thanks. There is still quite some work before I can make a PR.

• The changes to sd.cpp are relatively trivial but they do require some interface mods (I think it is better to make a switch for FreeU on or off).
• The ggml mods are more substantial, but still manageable. The problem is the two operators (FFT_FILTER and FREEU_BACKBONE) I added for FreeU are too specific and may not get accepted. I see there are some ggml extensions in sd.cpp but I don't see how cuda backends can be added. I am seeking your advice.
• The two additional operators are only implemented for GPU backend and I have to work on their CPU counterparts.

@FSSRepo, I can verify that the two operators are functioning correctly as I have test cases for both. But I don't think I can replicate results with sd-webui or with the implementation from the original author's repository.

[bssrdf]

I gave a try but I am not sure if implemented correctly.

@bssrdf Do you mean in sd.cpp? I checked your fork, and I don't see anything related.

Yes. Please check out add-freeu-support branch of my sd.cpp fork as well as add-freeu-support branch of my ggml fork. Most of the work is done in ggml. Sorry, code is a bit messy now and I will do a clean up before the PR.

[leejet]

I gave a try but I am not sure if implemented correctly.

@bssrdf Do you mean in sd.cpp? I checked your fork, and I don't see anything related.

Yes. Please check out add-freeu-support branch of my sd.cpp fork as well as add-freeu-support branch of my ggml fork. Most of the work is done in ggml. Sorry, code is a bit messy now and I will do a clean up before the PR.

Great job! Can you create a PR for both ggml and sd.cpp?

[lpn256]

FreeU_v2 seems to be a paper update: https://twitter.com/scy994/status/1714499568573039102

[FSSRepo]

I gave a try but I am not sure if implemented correctly.

@bssrdf Do you mean in sd.cpp? I checked your fork, and I don't see anything related.

[FSSRepo]

@bssrdf Great job, to implement that functionality in sd.cpp for obvious reasons, it must also be compatible with the CPU backend. I don't know much about how to implement the FFT optimally, so I can't help you with that. I'm also interested to see if you could replicate those results with sd-webui or with the implementation from the repository to check if they match or are similar. Additionally, it would be helpful to determine if certain values passed to the Fourier_filter function are the same as those you receive with your implementation.

def Fourier_filter(x, threshold, scale):
    # FFT
    x_freq = fft.fftn(x, dim=(-2, -1))
    x_freq = fft.fftshift(x_freq, dim=(-2, -1))
    
    B, C, H, W = x_freq.shape
    mask = torch.ones((B, C, H, W)).cuda() 

    crow, ccol = H // 2, W //2
    mask[..., crow - threshold:crow + threshold, ccol - threshold:ccol + threshold] = scale
    x_freq = x_freq * mask

    # IFFT
    x_freq = fft.ifftshift(x_freq, dim=(-2, -1))
    x_filtered = fft.ifftn(x_freq, dim=(-2, -1)).real