Please use this topic to ask any question about the course (even if you think they are stupid, we really love stupid questions).
Questions with lots of likes will be answered live. If you see a question you’d really like to be answered, click on the heart under it to like it. Don’t post duplicate questions if possible. They will be removed by the moderators (don’t be mad if this happens to you, it’s nothing personal). If your question is not answered during the course, it will be answered on this topic after the lesson.
This topic will be very crowded, so please refrain from getting too much off track (create another topic for this) and please like the post of someone who replied to you instead of replying with thank you.
Hi Jeremy, I just started to watch and take notes on part 2 lectures in 2019. I wonder how would 2022 part 2 be different from part 2 of 2019 apart from adding stable diffusion to the course. What are the other differences or new things should we expect this year?
It will follow a similar structure to the 2019 course but with lots of additions and increased depth in many sections. For details, you’ll have to watch the course!
I wonder what kind of compute is advised for this iteration of the course? Is it similar as for the first part, i.e., a collab instance, single (low-medium) GPU machine, Kaggle notebook? I’m mostly thinking about things like diffusion models, transformer-based NLP architectures, etc. If I am not mistaken, many of them became less computationally demanding during last few years.
I think here we might learn about stable diffusion and similar stuff. They requires a bit of GPU compute specially to train. Mostly VRAM size. Since colab has T4, it will work.
In the last few weeks training for dreambooth and similar got improved a lot too. Now we can train them on colab too.
Just browse the Stable Diffusion Reddit thread for latest improvements on this area.
How do we use real objects in the image generation process. Example, If i have a Specific Study Table, and want to create a living room with MYStudy Table, how should i go about it? I know that Dreambooth exists, but it is still not capturing 100% Fidelity, it still kind of distorts the product. How do I go about creating one, where there is no distortion of object in the generated image.
You could use a technique referred to as outpainting where you ask the image generator only to apply its generation to areas outside of the object, this is assuming you don’t want ANY changes to the Table.
There are also parameters for the image generation process (i.e. using img2img) where you can optimize to hold the same structure or content of the original object/image with minor changes to the style for example.
Hi, I’m wondering if there is any info on environment setup that we should be doing before the lesson? I was intending to run notebooks locally. I didn’t do part 1 of this course so it’s possible I’ve missed the instructions on this. Is there a requirements.txt for the course?
Generally folks have been running stuff on Kaggle, Colab, or Paperspace. For running locally, you’ll need PyTorch, fastai, and the HuggingFace libs installed all with CUDA working.
You might want to check out the docker containers by seeme.ai or paperspace. For Part 1 I used the paperspace container, but I see that the seeme.ai container is newer (pushed a month ago vs paperspace 3 months ago) and smaller than the paperspace container.
Hi Amir
I had a similar confusion. I believe you are all set if you are here and can posts messages into this section of course forum. Check ** About the Part 2 2022 course** for the details on how to get to the live stream.
