The intuitive derivation part is really beautiful. I was finding it hard to grasp the concept previously with just the paper derivation but this makes so much sense. And to me personally, the intuitive derivation as well as the final code (specially the simple sampling part) seems to be even more helpful if I think of the application with different interpolation methods for different noise.
I love these videos. I always struggle a bit to get to fully understand these concepts by reading the papers, but how you explain it really makes it stick.
I see. Very nice. I remember reading the paper and not understanding how the x_1 data was matched with the x_0 noise. I did suspect that it was just random, but thought that wouldn't work. The way you explain it makes it seem more reasonable that it would.
Really great video. I implemented this in pytorch just based on my memory of your intuitive description and it...just worked? Transformed 2d gaussian noise to a "ring"-shaped distribution. Pretty awesome :)
I love the way you showed how to move from the FM to the CFM objective! In Yaron Lipman’s talk on the topic, he simply mentioned that the gradients coincide, but your derivation helps a lot with the understanding! As a follow-up, I would be very interested in a video on conditioning of FM models with CFG, like it was done in the paper “Scaling Rectified Flow Transformers for High-Resolution Image Synthesis” (Esser et al.). Keep up the great work!
Really nice series on Diffusion models. Thanks for doing them. I'd like to understand the ideas behind single step Generation models more, such as Consistency Models or Rectified Flow Models
The authors of this paper arrived at the same approach, but derived it rather by intuition: "Iterative α-(de)Blending: a Minimalist Deterministic Diffusion Model" (arxiv 2305.03486).
Many of the methods involve image generation for goal setting or as part of the world model, there is lots of research involving multimodal models and about learning compressed latent spaces. Would fit well with your other videos. Something like dreamer v3 would be cool, just as an example. Showing a computer program learning to mine diamonds in Minecraft is also very spectacular
You have made some incredible videos, I come from a Mathematical Physics background and I really appreciate your mathematical exposition of these machine learning papers on image generation! Thank you very much!!! If I could suggest another topic to make a video I would like to suggest a video on Nvidia''s new open source model SANA that affirms to be more efficient than the majority of image synthesis generative models. They Talk about methos such as DiT with linear attention and a Flow-DPM-Solver which I don't know if these are the same as Flow Matching.
The intuitive derivation part is really beautiful. I was finding it hard to grasp the concept previously with just the paper derivation but this makes so much sense. And to me personally, the intuitive derivation as well as the final code (specially the simple sampling part) seems to be even more helpful if I think of the application with different interpolation methods for different noise.
😃 Thank you for making this. I’ve been looking for an explainer like this for ages. Always learn new nuggets of info I missed from papers.
amazing video. the efforts you spend to really understand and the creativity you have to make things simple to teach really shows.
@@gnorts_mr_alien thank you :3
Finally! I've been waiting for this video. Just started watching it, but I know it is going to be great! Thank you :)
I love these videos. I always struggle a bit to get to fully understand these concepts by reading the papers, but how you explain it really makes it stick.
Awesome video! Impressive how you can break it down to such a simple idea 👍
I see. Very nice. I remember reading the paper and not understanding how the x_1 data was matched with the x_0 noise. I did suspect that it was just random, but thought that wouldn't work. The way you explain it makes it seem more reasonable that it would.
Really great video. I implemented this in pytorch just based on my memory of your intuitive description and it...just worked? Transformed 2d gaussian noise to a "ring"-shaped distribution. Pretty awesome :)
@@jojodi oh that is awesome
I love the way you showed how to move from the FM to the CFM objective! In Yaron Lipman’s talk on the topic, he simply mentioned that the gradients coincide, but your derivation helps a lot with the understanding! As a follow-up, I would be very interested in a video on conditioning of FM models with CFG, like it was done in the paper “Scaling Rectified Flow Transformers for High-Resolution Image Synthesis” (Esser et al.). Keep up the great work!
Really nice series on Diffusion models. Thanks for doing them. I'd like to understand the ideas behind single step Generation models more, such as Consistency Models or Rectified Flow Models
@@audioentropy6242 thank you! My idea for the next video is distillation which kind of aligns with what you suggested
@@outliier Nice, looking forward to that. Will be very helpful for me🙏
Great video as usual. Keep it up!
The authors of this paper arrived at the same approach, but derived it rather by intuition: "Iterative α-(de)Blending: a Minimalist Deterministic Diffusion Model" (arxiv 2305.03486).
Extraordinary video.
You always give me so good pointers and intuitions to go on and read my papers. Thanks so much for your work!
@@zuchti5699 thank you! Could you point to me to an exact timestamp where this could be improved?
Great video! At 14:50 you mentioned it's canonical transformation for Gaussian, does this still holds for non-gaussian distributions?
@@seehow-ai yea you can also use other distributions
Nice video's! I'd like to see RL video's, especially model based RL
Many of the methods involve image generation for goal setting or as part of the world model, there is lots of research involving multimodal models and about learning compressed latent spaces. Would fit well with your other videos. Something like dreamer v3 would be cool, just as an example. Showing a computer program learning to mine diamonds in Minecraft is also very spectacular
amazing video bro
hi , at around 9:26 is q(x1) the probability density of the marginalization variable x1 ? thanks
Yea and technically this could also just be written as p_1(x_1) as far as I understand this
But this is how the paper did this too so I kept this part
You have made some incredible videos, I come from a Mathematical Physics background and I really appreciate your mathematical exposition of these machine learning papers on image generation! Thank you very much!!! If I could suggest another topic to make a video I would like to suggest a video on Nvidia''s new open source model SANA that affirms to be more efficient than the majority of image synthesis generative models. They Talk about methos such as DiT with linear attention and a Flow-DPM-Solver which I don't know if these are the same as Flow Matching.
Thanks a lot
Thank you for the nice video! I always found diffusions models to be packaged into way too much math compared to the actual implementation.
another banger video
You're awesome! I would be curious to know whether you're a PhD, in industry, or self-taught?
@@SN-uc3vr mostly self taught, studied a bachelor on AI in Germany and now I work in the industry at Luma AI :3