Congrats mate for 1000 subs! Cool informative video, love the deailing. Just one question the text desciptions next to every image in the bins for the traing data is it done manually? -I mean it is the prompt for the image right and saved right?--
@FiveBelowFiveUK
5 ай бұрын
Yes and yes. I recommend using a "caption root" approach. I sort my images into bins by similarity using a few aspect ratios and scales for each group, the name for the folder should be the "start of the prompt" this becomes a phrase trigger. I just renamed the contents of the folder after the folder name, then used the script to create all the text pairs. After this we can use an autotagger merging solution if we want to enhance the captions further. (civitai has a good example)
@Ysfsyhn
5 ай бұрын
Hey there! I've been a follower of your channel for a while now, and I've really enjoyed the tutorials and information you provide. I've noticed that you often focus on surrealistic images and outputs, which is great and really creative! However, I was wondering if you might consider incorporating some content on realistic image generation, particularly using Loras for human characters. I think it would be a valuable addition to your channel and would love to see your take on it. Thanks for all the great content, and keep up the fantastic work!
@FiveBelowFiveUK
5 ай бұрын
it shall be added to the list! Currently we are following a meandering path through my comfyUI crash course, with a few new models and features being showcased. I have a soft spot for abstract surrealism, certainly true. I like to portray my own work often as a "your art goes here" to encourage people to get training :) however i've done a lot of stuff in photorealism so i'm glad to revive some old datasets and show how i approach that in future
@FiveBelowFiveUK
5 ай бұрын
Now Let's see Paul Allen's Lora: civitai.com/models/405640?modelVersionId=452240
@Ysfsyhn
5 ай бұрын
@@FiveBelowFiveUK Your tutorials have been incredibly insightful and inspiring. I thank you for the effort and dedication you put into your content. It's been a joy learning from you!
@ADELTUF
5 ай бұрын
so it is possible to train it using video inputs?
@FiveBelowFiveUK
5 ай бұрын
yes, simply dump the frames to images, however you might only use keyframes, because there will be lots of them, i used an Automatic1111 extension to do this easily in UI, however you can use FFMPEG for this too. You dump the frames and then tag them as you would any image model dataset. There are ways to use video frame sequences to train other models, like motion loras with animateDiff, however this guide is only dealing with single frame image generation loras :D
@ADELTUF
5 ай бұрын
@@FiveBelowFiveUK it would be nice to see you do a video about this topic, the AI is able to output realistic images and videos, so it is likely possible to dump some videos to train it, and make it cook similar videos just by changing the prompt.
Пікірлер: 9