--- v4 ---
Improved fantasy awareness
General Improvements
Minor bug fixes
--- v3 ---
Improved skin textures
Better environment adherence
Eye Improvements
Shadow and Light improvements.
--- v2 ---
Improved prompt adherence (especially for non-human requests)
Better animal generation
More refined textures
Eye improvements
Lighting Improvements
--- v1 ---
QuadPipe is an incredibly responsive photo realistic model. It maintains native tagging while enhancing the quality of the output to a new level. QuadPipe handles text about as well as possible for SDXL and seems determined to get it close.
This is my best work. I've never created a model that performs better than this one.
I will continue to update this, though on a slower scale, to make sure that things are very refined before release. As always, let me know if you find any bugs, and I'll incorporate your discoveries into my next build. Thank you all for your support, I couldn't have created this model without your faith in me.
Check my prompts to learn about settings etc... but in general this is what I was using when I created the demos:
768x1152
CFG 5.5-9.5
4x_foolhardy_Remacri (upscaler)
30 hires steps
0.27 Denoising Strength
Upscale by 1.5
Sampler: DPM++ 2M SDE Heun
Schedule: Karras
Sampling Steps: 50
Description
FAQ
Comments (20)
This is a seriously nice model! If I may ask, what sort of dataset was it trained on?
I do a lot of iterative and incremental training, usually in the form of dozens of smaller, specifically trained LoRAs that are merged in. I started with the original SDXL and a single fine tuning, then I look for areas to improve and build datasets around those areas through focused generations and training. Depending on the concept it could be anywhere from 40-300 images per LoRA dataset.
Thank you for the kind words!
Oh duh! Forgot to actually answer your question!, all of the images for the datasets are generated by the most recent version of the model - usually a few hundred for every 1 or 2 selected for the dataset.
@QuadPipe Very cool, thanks. Sort of a self-improving model then. :)
I've actually been using a couple of other models that have specific strengths as the source for my datasets too. I'm definitely adding in this one as a source of images! Thanks for the model, great work!
@TrueToLife_Fauxto Thanks so much! I think that's the smart move. Too many moving pieces to have a single model for everything, I have a bunch of models I maintain for that reason... they all have their "thing" their good at.
fyi, I'm working on a hands update for next v now. Hands always seem to lose some integrity as the models advance, and I have to dial them back in for some reason.
Thanks for the new update! It looks great and I can't wait to try it out! There's something about SDXL that really lacks in Flux and that's the amazing textures, creativity and tonal variety. Most Flux models produce boring sterile plastic-looking surfaces and objects look like CGI, so I still love using SDXL even though I can run Flux fine on my RTX 3090. Keep it up and Happy New Year! :)
ty ty, I feel the same way. I'm experimenting with 3.5 now and like it better, but we're a way off from that being useful.
@QuadPipe Yes, SD3.5 looks very promising. Good to hear you're experimenting with it.
BTW, not sure if you already saw these, but there were some interesting articles on Civitai about improving the training capabilities of SD3.5 (and Flux) thanks to some new interesting findings:
1. Unlocking SD3.5's Potential: Fixing t5 max length for Better Training Results
2. The Hidden Flaw in AI Training: How Misaligned Text Encoders Impact Your Models
3. Revolutionizing CFG: Unlocking Flux Distilled Models with Advanced Guidance Algorithms
@mmdd2543 Awesome, I'll take a look. I'm fortunate to have a 4090 here, but even with that, I have to offload some training to the cloud periodically. The 3.5 has been a little tricky for me so far, so I'm definitely nowhere near competent with it yet.
Awesome model, very clear and accurate..
if I may inquire.. which tool was used for training and on what settings? Ive been trying to do a fine tune of SDXL and all my attempts failed miserably using Onetrainer
I use Kohya_ss, and the settings depend on what I'm training. Sometimes, I use the trainer here when my machine has too much going on, and I typically keep it to stock settings, except that I set it to 20 epochs and ensure that my tagging is thorough. I also make sure I have a minimum of 3000 steps per LoRA.
But that's my real key. I create dozens of small concept and refinement-based LoRAs and merge them into the last version of my model to improve it. It all starts with the base SDXL and then each version is refined after that. You can merge the LoRAs into your model with Kohya_ss as well.
https://github.com/bmaltais/kohya_ss
@QuadPipe Oh so you create LORAs and go merging them to the base SDXL instead of finetune.
Haven't tried that, but prolly works better than the Finetunes, that seems to be extremely hard. Thanks for the info!
@PabloFG When I first started experimenting with SD models (which feels like a million years ago now) I tried to strong-arm the models with traditional fine-tuning but I didn't have the kind of control I wanted. So the LoRA approach has been good for me because it's faster and I can remerge at a different weight if I accidentally overpower the merge or under-power it.
I guess I should add that when I first started the LoRA approach, there were a bunch of people who disagreed with my approach - and they might still. But the results have been solid, so I might be prepared for some naysayers until you can just show them what you've done. They were out of control trolls honestly, but I haven't heard from them since I pushed out the work.
@QuadPipe Ah in my case, TBH, I don't even have much of an option with Onetrainer: the Finetune fails every time, I tried every single combination and LR. Either doesnt move or deep-fries.
So this sounds like an approachable approximation. In any case I m doing a Lora on R256 today and the result is pretty close to what the finetune looks (until deepfried at least) so arguably a Lora at 256 is almost almost a fin-tune on itself, without the inherent basewiehts burn issue.
Whatever gets the job done.. Ill do a proper test soon.
@QuadPipe This is amazing, remarkable even, I never even considered this a possibility nor have I heard of anyone else taking such an approach. Bravo to you for thinking outside the inference box. May I ask if you have shared details on your LoRAs for this process; specifically the general number of images you're working with and or configs? Is that something you're willing to discuss publicly or privately if say, paying for your time is the better approach to a busy lifestyle. :) Thank you for all that you've contributed, I remember seeing your first Pony model, the first I heard of you and now, wow... many a models. Kudos!
@moocoop first, thank you for the kind words!
I'm happy to share things that have worked for me - I think we're in that stage where there are a lot of paths open.
I've been considering doing an article or article series on it, so this is a good time for me to sit down and get some thoughts out.
This model has great prompt adherence. Thank you for sharing with us. Would it be possible to share the tags/prompts it was trained on with us? It makes it easier to figure out which words the model resonates with, sometimes an adjective will barely work but a synonym will work way better, likely cause the model was trained on that instead.
It's not as simple as that with my models, but I try to retain core tags. My approach for a model begins with the base model (in this case SDXL) and then I create man individual LoRAs to refine it until I get it to a point where I'm ready to release a version. Then for the next version, I start with the last version of the model and continue to refine - typically along areas that are noticed in the community or areas that I notice that need improvement. So it can take dozens of conceptual and styled LoRAs merged into the model before release and those settings vary from LoRA to LoRA depending on what the subject matter is, how saturated those tags were in the core Laion 5b dataset etc... Sometimes you just need to hammer a concept in and other times it gracefully will pick it up.
I used to do training directly on the model but I found I didn't have as much control as I have with the LoRA merges.
Thank you for the kind words, Prompt Adherence was a big priority for me with this one, and I tried to keep it as close to the original SDXL tagging as I could.
Fantastic.