Sliders Supercharged: Lora16, CLIP improvements, and More! π¨β‘
Added 2024-07-12 20:21:49 +0000 UTCπ¨β¨ Hey, AI art enthusiasts! Melody here with some slider-tastic news! π
We've been cooking up some seriously cool upgrades in the NTC Slider Factory kitchen, and we can't wait for you to taste test them! π
π₯ First up, we've upgraded lora to rank 16 β it's like upgrading from regular fries to truffle fries. Slightly bigger portion, but oh boy, the flavor! πβ¨
π€ Next, we've taught our CLIP a new dance routine. It's got some fancy new training moves that'll make your art pop! ππΊ
For our code-savvy friends, peek behind the curtain here: https://github.com/ntc-ai/sliders-conceptmod/blob/main/conceptmod/notrigger/train_notrigger.py
π§ Last but not least, we've fine-tuned our merges like a DJ perfecting their mix. Time to crank up the volume on your creativity! π§π
We're always pushing to be the best in class, because you deserve nothing less. So slip on those creative shoes and let's make some AI masterpieces! ππ¨
Got thoughts? We're all ears! Drop us a comment and let us know what you think about these juicy updates.
Stay creative, stay awesome, and keep sliding! ππ
Comments
Thanks you so so much! I will try this out!
Timothy Kautz
2024-07-19 03:20:56 +0000 UTCHi, thanks for the kind words and your support! Regarding training a no-trigger LoRA, let me clarify: In SDXL the notrigger is actually three trained networks, clip1 clip2 then the unet. You do have to combine them. Here's a script to do it https://github.com/ntc-ai/sliders-conceptmod/blob/main/conceptmod/notrigger/combine_loras.py Scaling from -1 to 1 this is an active area of research. Each network has to be adjusted separately and different terms train differently. The above script has a way to change the weights of each. Training notrigger will show the percentage in the loss the clip is from the target location (from the starting location). It's a bit confusing but what you are looking for is the distance percentage to go down (< 10% on sdxl is good, pony is higher). The readme definitely needs to be updated. Cool to see you training locally!
NTC
2024-07-18 21:14:02 +0000 UTCHi! I wanted to say thank you for all the recent updates! I am excited to try the Rank 16 LoRAs, though very sadden by the friction you have had to face fixing the ComfyUI / LyCORIS DoRA issues. I'll certain be sad to see DoRA support vanish from your training solutions and am contemplating using a patched version of ComfyUI. I do have a couple of questions that I hope you can help me with. Could you please help clear up my confusion in training a no-trigger LoRA? It seems like, when training a no-trigger version that it only trains the Clip Text Encoder - and that model output does not seem to work. The thought came to me that maybe we have to combine the text encoder with the unet of the regular training? Is this correct? Are there any other steps? How would you ensure that the unet is also scaled from -1 to 1? I would love to have some sort of example of how to train a complete no-trigger slider, combine them, or at least some tips or an updated readme. π I'm kinda desperate for a solution here. These are some really great updates though. Thanks again for your on-going support and empowering of the creative ai community!
Timothy Kautz
2024-07-18 07:59:50 +0000 UTC
