[Note: The author of this model - c0ur4ge - will release a more official version of a LoRA soon. Follow them for the update]
A LoRA extracted from the C4PACITOR model by c0ur4ge who deserve all the credit for this model.
Why I created this LoRA :
To use on CivitAI generation services, since base models are not supported yet.
To control the strength.
To combine or merge with other LoRAs and base (dev/schnell) models.
Description
FAQ
Comments (6)
Hey there! Creator of C4PACITOR here - just a heads up, while I don’t mind sharing this LORA, I do plan to release the “real” version of it in the next day or so and it definitely isn’t this large - it might be worth calling out how you’ve produced this. That said, I’ll be curious to compare this extract to the “official”! I appreciate the interest and credit. :)
Can you please also fix the link to the C4PACITOR base model as it's clearly been pasted from something else and doesn't lead to the actual model? Thanks!
Sorry, I was uploading a bunch of LoRAs and I messed the links up :p , Thanks for pointing that out. I'll add a note regarding the official version. I used SVD of the difference between C4pacitor and Dev to create this LoRA.
@nakif0968 Nice! I invite you try out the version I kicked out today! It’s about 1/5th the size and should deliver the punch you’re looking for. I’d be interested to compare them - I’m surprised at the size of the diff!
Where is this official C4PACITOR LoRA? I can't find it.
@civitai3463 I used to release versions of it but as time went on, it didn't prove very popular and we also moved toward a fuller finetuning process. That said, LoRA can still be distilled out of flux dev as a difference pretty easily and doing so at rank 16/32 would likely yield something pretty akin to what you're looking for. If there's appetite for any of the legacy LORA versions, I could be tempted to release 'em. :)
@c0ur4ge personally I'm always interested in LoRA over a full finetune, if the results can match. My use case is that I'm using the fp16 models chained together with various ADetailer-like approaches, and the LoRA approach lets me keep the same model in memory with "only" various combinations of adapters applied.
Plus (and this is a pretty far 2nd, but it's there), the disk space savings is pretty nice. :)
Guess I should mention, given the amount of entitled comments you get on your various uploads, I'm not one of those people. Just weighing in on your "If there's appetite [..]" comment! :) Thanks for everything you release.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.