KonPaEvo
KonpaEvo Mix is supposed to create a clean and vibrant base style. Artist and style tags should still work perfectly fine.
Use with CFG +-2.5 | range of 1.5-3.5 seems good, go higher for more vibrancy (and also brightness) (Comparison image CFG 1-6)
4S Lightning Lora @ 8 steps | 8S PCM Lora @ 16 steps
PAG scale @ 1.6 @ CFG ~2 (Perturbed Attention Guidance | thanks @fhaifhai)
Positive prompt masterpiece, best quality, very awa + optional artist/style tag(s) | very awa makes images more artistic
Negative prompt I personally use:lowres, worst quality, low quality, simple background, bad anatomy, bad hands, signature, text, watermark
Settings:
Sampler: Euler a beta | Deis beta | DPM++ 2M (SDE)
Steps: ~24 | 4S Lightning Lora @ 8 steps | 8S PCM Lora @ 16 steps
CFG: 1.5-3.5 | 3.5+ for more vibrancy (& brightness) | ~2 for PAG @ 1.6
v1.0 Merge recipe (simple | see "About this version" for subsequent release's recipes):
Base (NoobAI XL 1.0 EPS): noobEvo v1 by @advokat for style/vibrancy
KonbiniMix v1 by @Shiiro0 merged in ~40% OUT blocks for style (mainly outer blocks)
PaSanctuary v2.5 by @FallenIncursio merged in ~38% IN Blocks for what I think should be backgrounds/composition (except IN6 because IN6 sucked for this merge; introduced extremely heavy headlighting)
v1.0 Merge recipe "extended": https://ibb.co/jvmFSpP
Clip: 60% noobEvo, 40% PaSanctuary v2.5
Please include the merge recipe for any subsequent merges.
New versions on this model page will include the merge recipe in "About this version"
Note about CFG 1 and samplers:
SDE samplers like DPM++ 2M SDE help quite well with the paleness. DPM++ SDE however is twice as slow as euler so I recommend against using that one. Remember to use Karras scheduler because normal doesn't seem to work
If you want to use Euler A with CFG 1: Use Luna XL VAE or Hi-Res Fix with Hi-Res Fix CFG at 4-8 can bring back vibrancy to the image, itll be slower than Hi-Res fixing at cfg 1 but the initial gen will be faster, seems worth it to me if you don't want to use SDE samplers.
KonpaEvo name origin: Konbinimix Pasanactuary noobEvo (also somewhat intended: konpa, jap. コンパ, supposedly is a type of drinking party so it kinda fits like konbini, convenience store)
Description
First release.
Most showcase images were generated using CFG 5 before i knew that low CFG had rather positive impact on the model so use 1.5-3.5
Merge recipe (simple):
Base (NoobAI XL 1.0 EPS): noobEvo v1 by @advokat for style/vibrancy
KonbiniMix v1 by @Shiiro0 merged in ~40% OUT blocks for style (mainly outer blocks)
PaSanctuary v2.5 by @FallenIncursio merged in ~38% IN Blocks for what I think should be backgrounds/composition (except IN6 because IN6 sucked for this merge; introduced extremely heavy headlighting)
if this edit gives notifications to the users mentioned above then im sorry and ill remove them but please message me if that is the case
"extended": https://ibb.co/jvmFSpP
Clip: 60% noobEvo, 40% PaSanctuary
FAQ
Comments (16)
Such an amazing model, well done! CFG 5 gave me too intense colors in many cases but when I reduced the CFG to 1, the default style changed dramatically but the results were still amazing. I can highly recommend trying this model with both CFG 5 and CFG 1. Most models become far too pale at CFG 1 but this one is still bright and colorful enough :)
Damn, I really need to test around more with CFG, I didn't even consider anything below 3.5-4 or higher than 6 because I had the prejudice it'd just be bad output. I never expected those results from CFG 1, they look unexpectedly (to me) good.
Also I read your article when I saw you (or I assume your bot idk) tested my (imo semi-failed) pony merge and I appreciate what you do, it's very helpful 👍
@DraconicDragon Thank you very much! Sadly it's not a bot, I have to do all this by hand ^^"
I tested your model a bit more and I think the sweet spot is at CFG 1.5, although that is just my personal preference.
@lizardon1024 By hand is incredible work, and also I talked with someone and they reminded me that CFG 1 almost doubles inference speed which is incredible, i've just assumed that it works for flux and for some reason I've never tested around or considered anything other than "default" or "recommended" values for CFG so seeing this in action is like discovering something completely new to me lol and now i'll definitely have to try around a lot more
@DraconicDragon Yes I can confirm that CFG 1 speeds up the generation by about 70% in my case but it looks like it creates more errors (additional arms, fingers etc.) compared to CFG 1.5. I will therefore use your model at CFG 1.5 for now :)
@lizardon1024 yeah that might be due to the missing negative, or something else, idk AI enough
My sweet spot will probably be around 1.5-2.5 too depending on what I want to create. Thanks again for testing, I probably wouldn't have ever noticed/tried this myself
Tested 4 step lightning lora and 8 step PCM lora, both work. Doesn't work with CFG 1; If you want the speed boost choose either CFG 1 or lightning (lightning is better usually).
Comparison image: https://civitai.com/images/41161512
(Since they are loras you need to use them at double the steps they are made for, eg. 4 step lightning lora needs 8 steps; 8 step pcm lora needs 16 steps)
This model deserve a tip only because you made a cute fennec girl!
Thanks!
I've done some tests with CFG 1 and here are the main points so far:
DPM++ 2M SDE Karras or Euler A with Luna XL vae works best.
Should you use CFG 1 or Lightning?
- Entirely up to you. Idk the quality loss of Lightning compared to CFG 1 but 4 step lightning is slightly faster, no paleness issues and has negative prompt.
CFG 1 is nice to have a change of style.
FreeU also seems to massively increase vibrancy to CFG 3-4 likeness.
For comfy users: you can increase vibrancy artificially by using a LatentMultiply node and setting the value to ~1.2-1.35 before VAE decode
If you don't want to switch away from Euler A but still use CFG 1 and get more vibrant results I strongly suggest using the Luna XL vae and hi res fix with a CFG of ~5
It's a model I really like. Thank you. I like this model because it has a variety of styles.
Hi! Your model is amazing, thank you for it.
Could you please recommend HiRes fix upscaler?
I haven't been exploring upscalers a lot but personally I felt that these are good:
4x-NMKD-YandereNeo-XL - good for anime, its what I mainly use
Remacri - it's what CivitAI uses in the on-site generator, seems like a solid all-rounder, used it before NMKD YandereNeo
4x-NMKD-Siax-200k - the upscaler I use for realistic stuff, seems to retain detail good and doesn't "inflate" the detail like anime upscalers or unmodified RealESRGAN does
I can also recommend RealERSGAN_x4plus_anime! That's what I always use for anime, since it gives nice clean results especially in fiddly areas
It is not an exaggeration when I say this is the best NoobAI based checkpoint on Civit.AI. I highly, highly recommend giving it a try and playing around a bit with tags to find the style you enjoy.
thanks for sharing the merge receipe ! it's great that you have keept track of it. and in ComfyUI on top of it :) !
still my favorite fix of NoobAI. zero Lora issues when trained on NoobAI 0.5 (and even illustrious too but more cfg sensitive) !
the style is a bit neutral but VERY high quality and sharp. which is great to edit. controlnet and ipadapter works wonderfuly. i should share some sample when i have time.
Details
Available On (7 platforms)
Same model published on other platforms. May have additional downloads or version variants.