spent some time on 🇨🇦 day making a version that's just c++ and compiled kernels most of the effort was actually a tool not seen here.. but then I didn't use the tool since I ended up automating the flying

love to see it

I miss this kind of AI video

Hell Yeah

What did the tool end up doing then

You can steer the video as it's rendering. Also, DD is modified so basically all the settings are per-frame in a CSV file. You can edit it as it's rendering to change the prompt or other settings. Works very well, but requires extreme patience to actually steer based on output.

Will you be putting this on GitHub?

I'm aiming for a simple Windows exe first. This isn't what most people who've gotten into it want first, but that's part of the point.

Idc I will use it with Wine. I want random Ai like this. I like having the option to do this kind of insanity

This is awesome

Is there anything you'd recommend if someone was going to have this recoded? Might try.

"rewrite this in c++ make no mistakes" hehe. In reality there is some iteration. I probably asked for c++ + libtorch first. It used TorchScript first but it was slow. Ultimately it's AOT compiled kernels that use fp16 without quality degradation. Just don't let it be mediocre.

Thank you so much.

@yacineMTB This is that circa 2022 ai video nostalgia vibe

What made you rewrite it? Still love disco, best creative experience and exploration I've had with any image creation.

I knew from recent experience that I can simply ask Codex or Claude to rewrite the whole thing in C++ and it's done in no time 😅 I do need to give it some advice to do things well.. and getting it to compile the kernels in the best way required persistence, but it's easy.

I had a tool for flying the camera during rendering in late 2021 (for Pytti, then DD) but I never properly released it because: a) it was written in Python (hard to distribute), b) required dev packages for GPU use, c) most users used Colab (ToS violation when driven by my app)

A C++ rewrite ironically (similar to how native iOS apps are often easier to build and distribute than Java) is easier to distribute. I can just pack an executable with the models and prebuilt kernels and people can install it. Python on the other hand is weird.

I also almost always end up iterating faster when a project is native C/C++ rather than something that runs slower. Rather than need to profile and target certain slow points, I can just make it all good and fast and basically not worry about it. Also unintuitive.

I hope you're going to release it with all the improvements you're cooking up. DD is something for the museums looking back at this time :)

nostalgic, will you push a code?

My primary interest has always been making an "it just works" app for Windows users who aren't developers or anything. I'm making some good progress getting AOT compiled kernels for it working on Windows MSVC (which is apparently broken in torch in at least 3 ways) so maybe 😅

Is this pytti?

This is Disco Diffusion (rewritten yesterday in c++) in 3D animation mode The loop is really similar since it uses optical flow to generate the init for each subsequent frame (pioneered in Pytti), but it uses CLIP-guided diffusion for the generation (not CLIP RGB opt nor VQGAN)

so you automated the part you built the tool for