Coding Stable Diffusion from scratch in PyTorch

As per my understanding , the loss function of an auto encoder is the KL Divergence loss. I dont know if I missed it the video but , I cant figure out where we have added the loss function .

一开始听口音还以为是老印

What nobody seems to explain is why CLIP model is chosen to produce embeddings. Everyone mentions how it was used to match images to text, but how is this relevant at all if we are using our VAE encoder which has nothing to do with image encoder they've used in CLIP?

Really really good video. Can you create a video about tokenizer from scratch? Many thanks!

Thank you so much! the best stable diffusion video I found!!!

the most powerfull deep learning videos in the world are on this channel

This is extremely helpful, can you please also make content for score base diffusion models and other more complex scheduling algorithms like Ranga kutta scheduling? Thanks a lot for your efforts

Could you please make a video on how to train a stable diffusion model? e.g. how many images do we need to train it? what types of images should we collect?

谢谢你，总算清楚sampler和unet之间的关系了

Thanks!

Really great video for understanding stable diffusion in detail. Thanks a lot for your contribution

Amazing job my friend! I just got a job in ShenZhen China by learing it! Thank u so much mate. I hope u and ur family living a great in China :)

pre-trained weights not working with the code you have provided.

I wish the code size was larger to make it easier to read.

In the Original Stable Diffusion Process, are the encoder and decoder components trained independently from the Noise Prediction U- Net architecture and then utilized as pre-trained models, where the architecture looks like Pre-trained Encoder + Noise Prediction U- Net + Pre-Trained Decoder (Note here Noise Prediction U- Net is not related to Pre-trained Encoder / Decoder before training combined Stable Diffusion )? or Are the Encoder, Noise Predictor, and Decoder trained together as a unified system, where they collectively learn patterns from the training images?

I just discovered a great, wonderful, amazing, fantastic, gem channel 🎉🎉🎉

It's the best explaination ever!!!! Thank you!

ModuleNotFoundError: No module named 'pytorch_lightning'

Awesome, This is the best explanation!!!

Thank you!

jesus I have base knowledge of AI and Statistics but you made me understand quite a lot of things thanks to your vid

This is amazing video!! Great job!!!

讲的非常不错！❤

guoqing jie laojia😂chinese?vary good video, keep going,Thank you!

So u didnt train the unet?

this covers LoRA? can you make a video if not?

Thanks Dear For helping Us , you Video's are very helpful

Is this coding compatible with diffusers library? I fine-tuned the stable diffusion model for my dataset but I need a torch model for further changes but I couldn't capture the full dependencies from diffusers code.

Great Work! Could you make a tutorial for ControlNet?

almost karpathy level explanations, thank you!

excellent video, full of information

Is it possible to create Stable Doffusion alternative using Brian2 instead of PyTorch?

Thanks again for the video. This is my second time watching this video. I can't help but notice that in the original latent diffusion paper, they were using vqgan to compress image into latent. Is the choice of VAE just for convenience?

An extremely detailed video about diffusion. I have learned a lot. Thank you ❤❤❤

What about training, I could not find a a training file in your github as well.

Really appreciated, very informative.

By far best explanation ❤

Your code is so detailed and it runs on my enviorment just fine. Great job!!!👏

Thank you so much for this amazing work!

Great work. I love this so much. Which auto completion tool are you using in VScode btw?

Why does the VAE encoder not use an activation between the two Convolution layers? Don't we need a nonlinearity?

Subscribed ❤❤

Thanks for this informative explanation. I was wondering in the demo file you took a pretrained model but you already built the model by your own, why don't you use that one, if I want to use how to do that? and Would you tell me how it will works for image conditioning without using the Clip text prompt? Thank you.

fabulous! thank you very much!

This is the best explanation of latent diffusion models I've seen

Sir, if we want to pretrain a distilled version of stable diffusion how to do that.

Great video! Really well made and informative.
On the sidenote.
Is somebody familiar with the correct VAE loss function? As i understand it consists of reproduction loss and KL terms. How should KL term be summed up over channels and batches? Also is there normalization for batches and training set?

Thanks for your video..I learn a lot..may I check how can I put Lora into it ? Thanks