To begin with, English is not my mother tongue, so I apologize in advance if there is any mistake.
I think this checkpoint can be unstable sometimes.
The images (character faces) I posted with this model have been inpainted at least one time.
The images are use multiples embeddings in the negative prompts so may be different if you don't have these embeddings.
Sampling method: DPM++ 2M SDE Karras
Sampling steps: 30
Upscaler: ScuNET PSNR
Hires steps: 20
Denoising strengh: 0.51
Please, don't use restore face.
I don't have done a lot of testing, and i won't (well Diablo IV is coming XD).
Please, please, experiment and show me what my checkpoint is capable of !
Description
FAQ
Comments (19)
dont have time to test this yet but hey demo pic looks amazing
and it is very good after testing :)
@bullseyetroll Thank you !
not able to generate any of the images in the preview the model is missing in some of the images
My images are using some embeddings in the negative prompt, maybe it's the problem ? I've put more informations in the model description
Where did you manage to get this sampler from? "DPM++ 2M SDE Karras" Very interested to test this out. So far the results have been great with DDIM.
To be honest i don't really know, i updated SD and it appeared. Maybe it's one of the extensions ? I don't know, sorry. If someone knows, please, enlight us.
@overlord91525 yeah asking because I have mine with automatic updates on and that one hasn't popped up, hence asking. Very interesting. Let's see if someone knows the answer to this.
@RykeWollf13 I have it, but I don't know when it appeared. My launch script for webui has git pull at the start, so each time I have the newest version. Could be a very recent addition. Try going to the settings tab and saving, then restarting webui - maybe it needs a settings save to appear.
@sheevlord It was definitely an update, i was running a few sampler tests that I had found but because sd doesn't update the file if i made changes to it, i wasn't getting it. This model is awesome though so worth losing the others, @overlord91525 what VAE are you using by the way?
@RykeWollf13 I used ema pruned vae or vae-ft-mse-840000, they produce similar results
Hey bud, thanks for your work on this model but my colors are really washed out no matter what vae I used (or none at all). Mind sharing what VAE you used?
I used ema pruned vae or vae-ft-mse-840000, they produce similar results
Its the same for me. The colors are very muted.
@scesher Strange, can you post an exemple ? or copy past your full prompt with sampling method/steps, etc
@overlord91525 I tried the vae you told me (which i had tried before) and I did get results close to yours, it just seems model is prone to more toned down images, which is fine, as long as it's your stylistic choice. I overall am very happy with the results.
@StereoNostalgic218 I'm happy to read that =)
Don't hesitate to post your creations !
About negative embeddings:
verybadimagenegative_v1.3 seems to generate images with 3 arms and legs or something like that. Can also ignore your clothing-related positive prompts creating less SFW results.
bad-picture-chill-75v seems to have the same clothing-related problem but overall it's fine. Ng_deepnegative_v1_75t is kinda a good alternative.
Personally I didn't like the results I got with badhandv4 so I replaced it with badv5.
What I recommend so far is the use of Ng_deepnegative_v1_75t together with badv5 because the hands in this model are quite inconsistent without any help.
I don't have much experience with SD yet but I thought this information might be useful.
Thank you, i'll do some testing later