Delcos
HD-17 seems to make img2img and inpaint progressively darker/grey-ish
This is especially noticeable with infill as it will show the infill area as a darker. greyer splotch on the image. I have noticed this with another model of which I have forgotten the name. So maybe it's on my end, but HD-16 doesn't show this issue.
Just wanted to chime in that I'm also seeing this. The first image is the txt2img result and the second is the img2img result with nothing being regenerated. Same prompt in both cases. Also seeing these strange blueish spots on img2img outputs. Using the negative prompt didn't change anything. I checked and this issue isn't present in HD-16.
Good to see I'm not the only one. I'd also like to confirm that I'm seeing those blueish/purple spots
I'm really liking the images I've been able to generate from this model, it's great stuff.
I will say that I've also noticed the purple splotches and it's the main thing I find myself fighting against.
It seems to show up most with img2img and using the Highres. fix, but it also shows up more subtly in normal or higher resolution images.
Try using a VAE. Blue artifacts can be eliminated with other samplers/higher step counts.
This happens with me as well for some reason, this issue does not show up in the previous model
I've been getting good results with the VAE from Anything V3. It seems to have some issues with half precision and higher res images, but it works great in full precision mode (e.g. use --no-half-vae for stable-diffusion-webui). It seems to fix the purple splotches and gives the images a bit more color.
This happens with me as well for some reason, this issue does not show up in the previous model
This is a known issue that is being fixed, and should be improved in 19 which will be posted later tonight.
Same. I am running serveral anime checkpoints (NovelAI leak, Waifu Diffusion, This one), and all images produced from HD are grainy/grayish, and washed out in color. I tried to combine this with WD to get the best results, but I still get somewhat grainy and washed out results.
I'm sorry but, which Lora? And where do I find it? The only thing included in the distribution is the checkpoint and the vae, and I'm using both.
On Fri, Apr 7, 2023, 10:03 PM Delcos @.***> wrote:
Use the lora. Manually set it if needed.
— Reply to this email directly, view it on GitHub https://github.com/Delcos/Hentai-Diffusion/issues/17#issuecomment-1500762225, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADG7J2PVFI5DILNPOXE2RQ3XADBQBANCNFSM6AAAAAAS3374TM . You are receiving this because you commented.Message ID: @.***>
Roger. I dumped both in the same folder (models), but I'm not sure how to tell if the vae is being loaded with the checkpoint. I just renamed it to exactly match the checkpoint name (it was slightly different) but I couldn't tell if its being loaded or not.
On Fri, Apr 7, 2023 at 10:31 PM Delcos @.***> wrote:
Sorry. I meant vae. Make sure it's loading it.
— Reply to this email directly, view it on GitHub https://github.com/Delcos/Hentai-Diffusion/issues/17#issuecomment-1500767395, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADG7J2MDK2UTIHLCGFTT3ZDXADEXJANCNFSM6AAAAAAS3374TM . You are receiving this because you commented.Message ID: @.***>
If you're using the UI we provide or Webui by 1111 then under the settings tab there is an option to manually load VAEs. It is usually set to auto but if it isn't changing anything try setting it manually. it will also work with most other anime type models so you shouldn't need to change it back and forth.