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
@Bigboyblaziken just for you uwu
Use CFG++ samplers or euler ancestral at very low cfg like +-1. Non-ancestral or non-SDE samplers wont work with this. This model is still an EPS model.
Also I still think this model is broken to a degree lol
This is a test merge I made some time ago. I merged in the train difference of Noob 1.0 VPred onto KonpaEvo 1.1 just to see what the effect was. I didn't plan on uploading it but the user mentioned above wanted it so I thought why not. The original effect that I originally expected was that I could maybe get extra knowledge trained into 1.0 vpred onto an eps model - and well the result doesnt seem to be that i guess
Also feels like me making some test version for each normal version is gonna be a recurring thing
Merge recipe: https://ibb.co/zR3BYYK
FAQ
Comments (7)
I only started using Illustrious recently, and this became my go-to model straight away.
Unlike others, you don't need to use a specific prompt and settings, specify an artist or use any extra Loras. Just wonderful images right out of the box with a simple prompt.
Question: I have been trying to make Loras with Illustrious, but they have not been working out so well.
Should I be training them on the base Illustrious-XL model or NoobAI-XL? I'm kind of confused where the line is drawn and if we now categorize models as Illustrious or NoobAI.
I am still a noob to Illustrious... and to noob. A noob-noob if you will.
I can't give that much insight to that question but as I see it, LoRA compatibility is a bit weird, I trained a LoRA on illustrious but it seems to have close to no effect on NoobAI based models. I noticed some people train their LoRA on NoobAI Epsilon version 0.75 to make the LoRA compatible with both Illu and Noob.
I think, if you have the resources (like a PC you don't mind having to leave a lone for a couple of hours or if you have enough buzz) train a separate version for each NoobAI and Illustrious BUT then again, I haven't trained a LoRA on NoobAI Epsilon 1.1 and tested it with Illustrious yet.
It might also depend on what concept you are training, the concept I chose (Snow Angel) was completely absent in any model I've tried so that could also be a reason why it had no effect on Noob.
About when a model is Illustrious or when it is NobbAI - if its trained on NoobAI it's NoobAI in my book and people should mention it in the description and I feel CivitAI should add a category for Noob already😄 especially for the Vpred version. The e621 knowledge in NoobAI already just makes a huge difference for me compared to illustrious and people won't write it in the description if their merge/finetune has that knowledge.
Lastly: There's a lot of factors that can go into why your LoRA isn't working well like is the dataset tagged correctly, was it under-/overtrained, possibly bad dataset, or bad settings, probably even more;
And also what matters: What models did you test the LoRA with is also a big one because if you for example train a LoRA on base illustrious v0.1 on civitai, the preview images generated during training will usually look quite bad which is usually due to, imo, illustrious v0.1 just being "bad" (in a sense that you need to use style tags and quite a few negatives to get a stable result etc, noobai base is the same)
PS: the CivitAI Discord has a #training channel where you can ask for some advice from other people too
I've done a lot of yapping, I hope some of that was useful to you lol
@DraconicDragon Thanks, I might have a go at training on the NoobAI-XL model next time.
I had a whole reply typed out, but while doing some tests with the few Loras I've made so far, they semed to work about as well as the Pony versions.. so now I don't know if I just imagined it or what..
After testing 1180 IllustriousXL based models so far, this one (v1.1) at CFG 1.5 is still my favorite. Without an artist tag it still offers a beautiful style, great details and nice backgrounds. Even compared to other ILXL models (which can almost all create thousands of different characters without LoRAs), this model is especially capable of recreating even complex and unusual details. Thank so you much for this model!
Would it be more accurate to list the Base Model as Illustrious rather than SDXL1.0?
If you are searching for checkpoints and filter for Illustrious, KonpaEvo won't come up.
I found this out while trying to add KonpaEvo as a Recommended Resource for my Loras (had to change the Lora base to SDXL, then I could select KonpaEvo, then change back to Illustrious)
Apologies if I'm talking nonsense.. I'm still a bit confused about what's Illustrious vs Noob vs SDXL
No you are right, I just got a bit worked up about the recent stuff that's been happening so I acted on impulse, changing the base model type to sdxl in hopes that someday a noobai base model category will appear that i can change to.
I'll change the main version's (1.1, 1.0) base model type back to Illustrious for usability.
As for the difference between the base model types - They are all ultimately SDXL.
CivitAI has different types to give user more of an idea of what works together and what doesn't, I assume
As in: SDXL LoRA's likely won't work with Illustrious or Pony
NoobAI is a large scale finetune of Illustrious v0.1 with Natural language captions and updated Danbooru dataset up to ~27th october 2024 as well as the 2024 e621 dataset (if you don't generate furry stuff the e621 stuff doesn't matter to you probably)
Illustrious v0.1 is a finetune of SDXL (Kohaku XL zeta I believe) with an older but still 2024 danbooru dataset (and in my opinion it's undertrained but merges/finetunes of it do the job completely fine)