Krita AI Diffusion 1.17, issues I am Having
HTML-код
- Опубликовано: 7 сен 2024
- These are issues I am having in version 1.17,
How is having these issues, have you found a way to fix it. reinstalled four times, two systems and bother have the issue.
------------------------------------------------------------------------------------------------------------------------------------------------------------
those with IP face issues try downloading these and place in lora dir
huggingface.co...
huggingface.co...
------------------------------------------------------------------------------------------------------------------------------------------------------------
(Be sure to support Acly So he can keep working on AI Diffusion for Krita)
Remember without there work we would not have this.
github.com/spo...
THIS VIDEO is for issues I'm personally having with Ai diffusion, hopping on feedback or others experience.
UPDATE
I'm working out the issues there sampler based I have some new ones that are giving me renders like i had before.
seems it was comfy but the ability to make samplers I getting it looking great again.
i will make notes and do video Monday. and get these new samplers up then.
people reaching out and there feedback has helped massively
back up your pykrita before updating, many are not having this issue, few of the samplers I added help just little slower, each of the installation for myself have been ontop of 1.15 so i may need to delete and do a full fresh install, sadly internet is slow and its why I have not done this as yet.
i do believe today that the issue is more from changes in comfy effecting the renders to be less clean looking and different.
They updated Krita Ai, this fixes a problem in the presets. Where the models have already integrated the LORA LCM they do not require the addition of the LORA, but if a preset was created that did not require the LORA LCM, the preset did not work leading to the generation of a saturated image as it did not work correctly or in any case was not loaded.
From the log of the Krita ai site I believe the error was also caused by this: Don't auto-convert previous sampler settings to a preset that requires a Lora(fix)
Now when creating a preset you can indicate to use lcm with or without lcm lora.
These are two presets created by me:
"NATIVE LCM KARRAS": {
"sampler": "lcm",
"scheduler": "karras",
"steps": 10,
"cfg": 1.5
},
"LCM + LORA KARRAS": {
"sampler": "lcm",
"scheduler": "karras",
"lora": "lcm",
"steps": 10,
"cfg": 1.5
}
This way I can use normal models with LCM that require the LCM lora and models that already include it and don't need the loaded lora.
@@Archalternative just woke and just got the 1.17.1 definitely has improvements
do you know how to add turbo
"lora": "turbo",
"steps": 10,
"cfg": 1.5
can work it out kicks out error, got a turbo lora but might be wrong one.
I'll try out the two you have, what models are you using those with?
@@streamtabulous For those models that require the lora I think you have to load the lora manually and not from presets.
That's great! Learn from you! thank you
I'm working out the issues there sampler based I have some new ones that are giving me renders like i had before.
seems it was comfy but the ability to make samplers I getting it looking great again.
i will make notes and do video Monday.
people reaching out and there feedback has helped massively
I can confirm, there are a lot of issues in 1.17. I experienced the same blurry images on my RTX 3080 Ti. On my side, there is no "Euler a". I can see "Euler" only, which means "Euler normal". Over all, I am not happy with the release so far...
i thought i was going insane, was preparing to do a review and have the prompts i use to keep consistency and was wondering what happened then bounced between then to realize there is a issue, one system ok error but two and both updated separately.
hopefully gets sorted, but one person doing it and free, i can't imagine its easy on them.
I'm fine on 1.15, as long as when SDXL3 is out its good lol
I encountered a server execution error while using the face (ip-adapter) control layer, which is quite disappointing. Unfortunately, the plugin does not provide an option to reinstall that specific part(or do the files checking). The issue with the messes graphics could be due to the VAE (incorrect pairing or failure to load). This is kind of a big problem for the plugin because it automatically selects the VAE without allowing users select it in the profile page or know which one is being used.
huggingface.co/h94/IP-Adapter-FaceID/blob/main/ip-adapter-faceid-plusv2_sdxl_lora.safetensors
huggingface.co/h94/IP-Adapter-FaceID/blob/main/ip-adapter-faceid-plusv2_sd15_lora.safetensors
have you downloaded and put these in the lora dir.
only issues im having with ipface is my sampler issues that i believe im sorting out and will do update date video Monday
let me know if this fixes your issue
@@streamtabulous Thanks, I checked and the items are already there. Renew the file doesn't help. Sad. I may need to delete the ip-apapter related folder to see it let me to redown that part or not... However, I was able to resolve the VAE mismatched problem by watching your latest video. I edited the "vae" line in the style.json file for that checkpoint to : "vae": "sdxlVAE_sdxlVAE.safetensors".
@@shanshanshans5167
check if you have these if not download them
huggingface.co/h94/IP-Adapter-FaceID/tree/36ce7f96f0c76c1e8b100f10fa094a2ac48ea1c6
ip-adapter_sd15.safetensors
ip-adapter_sdxl_vit-h.safetensors
ip-adapter-faceid-plusv2_sd15.bin
ip-adapter-faceid-plusv2_sdxl.bin
located
C:\Userspc\AppData\Roaming\krita\pykrita\ai_diffusion\.server\ComfyUI\models\ipadapter
huggingface.co/h94/IP-Adapter/blob/main/sdxl_models/ip-adapter_sdxl_vit-h.safetensors
huggingface.co/h94/IP-Adapter/blob/main/models/ip-adapter_sd15.safetensors
@@shanshanshans5167 normally to add vae its configuration / styles
checkpoint configuration advanced
vae drop down
and then any vae you have or have downloaded you can select
I have the same problem on a 1070. What graphics card are you using? (Just to try to narrow the range of origins.) Mine works with 1.5 adapter for xl is loading, runnig to a certain point and then giving a dtype error message.
My ai genrative work very slow. Can you have any idea to speed up it..? It take too many time to generate
Hello Streamy. There are some more issues, i have now which i never had. I am using a 1070 and i would like to know if you encounter them on either or both of your systems:
All features involving ip adapter plus reveal the old dtype error message again: face, reference and the new style transfer feature - but only with sdxl models. With 1.5 it is working. Acly assumed it is an architecture problem of the 1070. Thats why i'd like to know, if the issues are occurring on your 3060 as well. I was very dissapointed as i encountered, them cause i 've been eafger to start with that feature, which i suggested to integrate. Furthermore, i only have euler a and euler, to choose from, next to a bunch of fantasy name like creative. No dpmpp-sde karras and so on at all. Of cause acly's work is still worth being appreciated as before. Thanks for this post.
odd face is not working on the rtx3060 there no error but just black most of the time and other's extremely artifices distorted and what looks like black statistic diffusion.
I don't have Euler at all on the 1070 system the 3060 has random ones showing like there is code for samplers elsewhere and its randomly reading it as comes and goes.
upscale does not work just resizes.
to many issues on 1.16 and 1.17 for myself on both systems. the RTX3060TI seems to have more probably due to using larger canvas.
tested face on gtx1070 and it works but black diffusion spots but much better than the RTX3060ti where face is definitely broken.
both tests identical both same SDXL model etc.
there something wrong, the fact two systems and the gtx1070 was better. i imagine code conflict i don't understand how there can be the different error and poor quality, maybe way it loads something skips a code or cache system.
very odd. its sadly unusable for quality alone for myself.
but no error on faceip on sdxl, also was twice as fast on the gtx1070 than the rtx3060ti that certainly is not right.
try downloading these and place in lora dir
huggingface.co/h94/IP-Adapter-FaceID/blob/main/ip-adapter-faceid-plusv2_sdxl_lora.safetensors
huggingface.co/h94/IP-Adapter-FaceID/blob/main/ip-adapter-faceid-plusv2_sd15_lora.safetensors
?whats this for
To try and gauge if others are having the issue and why my lack of videos.
Can't do much when my systems not working correctly.