@alcaitiff whoa. only 435 images??? thats very low for something like this.

impressive.

what'd you train with? AI-toolkit?

sorry - I'm gonna ask a buncha questions because I'm very impressed. this has been the best NSFW so far and my loras have struggled.

what's your learning rate? I'm assuming everything is captioned. . . I'd love if you could upload your training config yaml somewhere to use as reference. I'd like to learn from the master

@alcaitiff yeah I hear you. I've learnt that the hard way. - on that note - I found the new https://github.com/Tr1dae/HuiHui-Batch-Captioner a much better captioner than joycaption. Still need to fix by hand but it's better from the get go and more promptable. Check it out for future

120 thousand steps is wild but it works. what was your learning rate?

@alcaitiff is the variable rate a strategy you commonly use or did you just adapt as you saw samples process? I've never considered doing this

@alcaitiff what batch size?

@QualityControl It's a large reason why schedulers such as linear, rex, etc. are a thing to change the LR when training. You can get the broad strokes quickly and then lower the LR over time to get details in without destroying the model.

@alcaitiff about your prompting strategy. My attempts to do this (results not as good as this) I have noticed that (as you have pointed out) qwen wants very descriptive prompts, whereas many guides out there tell you to use keywords. With qwen, it seems its understanding of the overall composition and elements within it is much stronger, have you noticed this to be the case? Do you do both keywords and a fully descriptive prompt?

I agree. I just tried this lora on a Qwen image edit workflow and the output is just awful. No other lora has this problem. Very strange.

same here, does not work with Qwen edit model 😒

@alcaitiff That's a shame. It's a really amazing lora for standard Qwen image.

try lora power 0.70, lightning lora 4steps 1.0. It is unstable, but work if you set 3 steps. (checkpoint gguf 4q_k_m).

@Helljeth I was not able to replicate your success at all :( Based on the ksampler previews it seems to work fine for the first half of steps but then the second half undoes the edits and adds bad warping. This happens with and without lightning lora and with any number of steps.

Where is GET_CAP_ID set? I'd like to use the file reader but I"m getting a similar error. I found where to set the input file path

In my testing Qwen seems to do this more than other models when using any loras. not sure how to stop it yet.

@J1B I think it's because faces were probably included in the training data, and probably multiples of the same person, so it learned facial features as part of the lora's data

Probably more asian than western faces

@alcaitiff Not sure what that is, but it sounds like something I would want! 😊

@alcaitiff you're a legend!

@alcaitiff Might you have a presence on another site with a nude beach?

@alcaitiff Come to the dark side—we have cookies.

@alcaitiff other nsfw lora like https://civitai.com/models/1851673/mcnl-multi-concept-nsfw-lora-qwen-image work with edit, but they are not as good as yours

@alcaitiff yeah dude, i don't understand why, its a porn site wtf

They want the "metadata"? OMG. That would be from "real" photos? They want our "real" photos? That's weird. Mine are all AI generated, so there is no metadata like that.

@ArtPencil yes, that's whats appear when i try with a real photo.

"We couldn't detect valid metadata in this image.

Outputs based on this image must be PG, PG-13, or they will be blocked and you will not be refunded."

@ArtPencil "We couldn't detect valid metadata in this image.

Outputs based on this image must be PG, PG-13, or they will be blocked and you will not be refunded"

?????

@Azathoth786  They are going to lose a bunch of money again.

@ArtPencil do you not understand metadata in the context of AI? metadata means stuff like prompt, sampler, etc to prove it's really ai generated.

@crombobular It was working and then it stopped. So there must have been metadata. I can see it in properties. So, I think this site will just lose a bunch of money again not truly putting in the time to figure out how their system works in many scenarios. IMO.

@Azathoth786 So, basically they want the metadata from the actual generation. So, the image to vid is more like generated image to vid That's what we did not understand. Say we screenshot an AI image, it still won't work because the actual metadata from the generation is not there. So, basically, if you want real freedom, you have to go open source locally, or rent a GPU. I think all the sites want that metadata so they stay safe not generating real people. It makes sense now.

@crombobular OK, I get it now. My metadata was still based on a screen shot from a AI vid I made so the "image" metadata is not there to detect. Image to vid really means "generated" image to vid so civitai stays safe from real images of real people like celebrities. So, since they are basically still censored, why not just stay with banks? They never needed to go crypto. They can still require metadata.