I was not able to get Freyja to do anything I wanted, at least not with natural language prompts.

@snipsnapsnipsnap that's not surprising really, ranking checkpoints is very subjective no matter what. we need to find checkpoints that "speak our own language". i tried to be as generalized as possible but inevitably there will be big differences in user experience.

I tend to disagree as well. House of Freyja gives the skin an airbrushed look - you can even see it in the previews. Much less skin detail than this model

ive heard a checkpoint named realistic freedom was the best

@rime11 in my tests realistic freedom did well in everything except prompt obedience. it seems to produce great stuff but rarely what is actually asked for

@shapeshifter83 Did you try CinEro XL v1.3 Handsome? This version had a focus on anatomy correctness.

sounds interesting, do you have a link to that.

@homoludens i did download it as soon as it came out but I haven't had a chance to really try it out. planning to

@twistedlegs https://civitai.com/models/151568

@shapeshifter83 if you will make a merge with cinero use elemental merge and try to find out which elements affect hands first. This is the main problem currently. Other things can be easily fixed by some loras.

Maybe upload some of your results to the Freya page. There's nothing there that looks interesting at the moment.

Did you find out how to avoid the long torso?

@edwin0815 Long torsos can have to do with the size of your output, try changing the dimensions and total size (pixel count)

Can you explain how did you create inpainting model version please? 😁

@TimNekk SDXL has no inpainting model. So use Fooocus to get inpainting feature, it's far better than SD1.5 inpainting

So, I discovered some things, I failed to recreate some of the authors images because i am using auto1111 and they are using comfy, it seems comfy uses the cpu instead of the gpu to generate the seed noise and the way the 2 systems weigh words are also different so it seems that if you want to use this checkpoint you should try it out on comfyui

not all loras require trigger words. when a lora is merged with checkpoint, it means that what the trained lora had learnt was passed on to the main model. detailed eyes has no trigger word, so the lora will apply automatically.

@aiaicaptain Does that mean, if you didn't know what went into training the LoRA, and you used both Detailed eyes and this one, you might end up with funky results?

wha ya mean creepy?

@ebony3d The second version produces images with faces that looked way too fucking young. It was really fucking creepy and I would encourage CivitAI to remove it from the website. I believe the pedophiles on this website are finding more and more clever ways to get around filters, and version 2 of this model is one of them.

this is related to your sampler, which you've probably already discovered since that was 11 days ago.

And also, you're being fooled by the image preview window in A1111. It's actually borked generation from the start, but the image preview window is showing you something different. Remember, the image preview in A1111 DOES NOT sample from the ongoing image being generated; it's a rough approximation from a totally different (and far cheaper) system running concurrently to try to give you the illusion of a progress window for the main job.

@shapeshifter83 hello

but my image at the end is blue not during the process ! so i dont know what i should do as i stop using it at the moment cause of this

i will try again later but thanks anyway for the answer

@rafa66 your image is blue throughout the entire process, the preview window is tricking you. The preview window isn't real. It has nothing to do with the final step at all. Try a few different samplers until it stops doing it, or possibly ensure you're running a good VAE (not the SDXL original, something like the fp16 fix or Sharp Spectrum).

@shapeshifter83 ok i'll try thanks

@shapeshifter83 That's an interesting insight. I suppose if that's the case in A1111, it's probably also the case in other packages like comfyui and fooocus, right? And here was I thinking I was watching the actual progress of the sampler.

@Janos_Thanatos I honestly have no idea if that's the case for other platforms, I only know A1111 personally.

It's the VAE, the other guy had no idea what he was talking about unfortunately. Hopefully you figured it out by now.

@ShencilLizard am i the "other guy that has no idea what he's talking about"? He didn't mention blue dots, nor did he mention NaN errors or completely black images, which is why I focused on getting him to stop trusting the image preview instead of immediately calling it a VAE issue. Even then, I did mention possible VAE issues in my second reply.

So what exactly is it that makes me "not know what I'm talking about"? You just saw him mention blue in his images and went straight to "other guy is an idiot, it's VAE", but he didn't actually mention orbs or dots, and have you ever seen or heard of a VAE issue causing a "completely blue and ugly" image? I haven't.