Phoenix by Arteiaman SDXL TURBO V1.0
Trained with the new turbo version of SDXL.
CFG:
works well with cfg 1.5-2.2 You can try others.
Samplers:
SDNEXT: Sampler DPM SDE and Karras
A1111: sampler DPM++ SDE Karras
If you want to use Euler A you must increase the steps to 7 or more.
Please do not use DPM++ 2M karras
Steps:
4 to 7 DPM++ SDE Karras (strongly recommended)
7 or more Euler A
Phoenix by Arteiaman V1.0
This is my first model and I have decided to publish it now that it is mature and sufficiently tested (in fact, it is the one I use regularly). It is a merge of several models and loras exquisitely selected from a multitude of models that I have been testing over the last few months. It has been a hard work of testing and testing and testing until I got this model which I consider to be one of the best for almost all purposes.
I think it is a unique model and I will have a hard time improving it but it is not for me to say so. I encourage you to try it and leave your reviews and pictures. I accept all criticism.
As I said before, it is a multipurpose model and works well with almost everything. It generates non-explicit NSFW content.
works well with cfg 5-8
no refiner needed.
Vae included
You can follow me on instagram @arteiaman and twitter(X) @arteiaman:
If you like my model you can invite me for a coffee
Feel free to post images with this model and let's check how well it works. I thank in advance all those who download the model and post comments with their criticisms that will serve to improve it in the future.
Description
Phoenix by Arteiaman SDXL TURBO version.
Trained with the new turbo version of SDXL.
CFG:
works well with cfg 1.5-2.2 You can try others.
Samplers:
SDNEXT: Sampler DPM SDE and Karras
A1111: sampler DPM++ SDE Karras
If you want to use Euler A you must increase the steps to 7 or more.
Please do not use DPM++ 2M karras
Steps:
4 to 7 DPM++ SDE Karras (strongly recommended)
7 or more Euler A
FAQ
Comments (17)
IT is not SDXL TURBO
The worst is people have false assumptions seeing name TURBO as is NOT model using technique ADD which was used under SDXL TURDO model.... where 1 step is fine.
It sounds like false advertisement... I do not like it at all.
I don't know how you are doing to get SDXL turbo to work with two steps because all the tests I have done, all the images are a disaster. Also, SDXL turbo only works with 512x512 so you lose the gain of those two steps when scaling the image. My model (and other turbo models) has a vastly superior quality to the original SDXL turbo and also working at 1024x1024 or more as you can see in my test images. Do the test with the configuration of my images in original SDXL turbo and you will see how much you laugh when you see the results.
@arteiaman REAL TURBO model actually woks fine with 1 step and is designed to other things like prototype and later for upscaling. That model is even works great on CPU .. image is generated in 1 SECOND .. GPU of corse something around 0.1s!
I wonder why people are not tunning real TURBO model.
Those models should be named differently not TURBO.
@mirek190 Yes, images in one second with hardly any detail and a lot of artifacts that then have to be scaled to look halfway good. Our models make good images in 4 steps without scaling, so our models are actually faster. Stability ai generates a poorly working base model but Stability knows this and expects the community to improve it. For example, the base model without turbo, generates very poor quality images but Stability knows that the community will tune this model to make it much better. This is how it works. The turbo model without fine tuning is useless. What we do is to add the base model to already tuned models and you can see that the images that our models get are impossible to get with the base model. If you think that our models should not carry the turbo surname maybe you are right, but in practice no matter how they are called, they are still as fast or faster than the base turbo model and with much more quality.
@arteiaman I understand but still ... real TURBO model is something different anyway... I wish someone could finetune real TURBO model :)
@mirek190 When you merge models like let's say juggarnaut with turbo xl does that mean you are making the turbo worse/slower? How does one actually 'finetune' a turbo model? Is it the same as finetunning a SDXL model?
@bigTiddyLover Normal SDXL model generating pictures with one step? No
@mirek190 No I meant if we merge normal SDXL with Turbo. Does the distillation come from another architecture or from the weights themselve? Because if it's from the weights, then that means merging with another normal model will reduce the speed.
@bigTiddyLover I guess I haven't looked too hard, but I haven't found anything on this subject, personally.
So, I took two Turbo models, merged them together; that worked, don't know why I was surprised.
Then I took that "Turbo-to-Turbo"-merged model, and merged it with one of my "Regular-SDXL" models; nothing special, just some merged bit of madness I had laying around. I felt much more justified by my surprise when this one actually finished without any error messages.
Using the ComfyUI 'realtime' workflow, that "Turbo-to-RegularSDXL"-merged model generates a decently coherent image by step 7 and is usually pretty clear by step 10 - which is about the same as my favorite Turbo model I've been playing with so far. The RegularSDXL model (when not merged with Turbo) is nowhere near that, with a blurry and incoherent image even at ComfyUI's SDTurboScheduler's maximum of 10 steps.
It appears that merging in this way somewhat alters 'which scheduler it prefers', might possibly raise the 'optimal number of steps', and blablabla. I'm not sure how or why or what I am doing....
but it appears that, yes, merging an SDXLTurbo model with a RegularSDXL model, appears to at least partly convert it into a Turbo model. The minimum number of required steps drops substantially.
I do not know (though I am extremely curious to find out) what sort of hoops I'd have to jump through to "seriously-turbo-ify" a merged RegularSDXL model / why it doesn't work as well at, say, 5 steps as the actual Turbo models I was using / what I'm doing (wrong/right), etc.
(I also tried to come up with a basic answer to your question of "is it the architecture or the weights" [because I am just as curious, myself] by changing which model my Merge Board was told to 'Copy Config From', but the answer is confusing: both versions are acting like Turbo models now. My brain tells me that it must be the "architecture"[config]; but my results say "🤷♂️")
@blncd Great answer! I came up with similar conclusions, but it is definitely nice for someone to confirm. Thanks!
@mirek190 This model can generate high quality images in 5 steps, which would normally require 20. To me, that's Turbo. As others have said, what's the point of generate garbage in 1 or 2 steps?
Maybe someday someone will figure out a way to create a fine-tuned model that can generate similar high quality images in 1 or 2 steps. Until then, I will be happily using this model 👍😁
@NowhereManGoes Do not care what you say.
Turbo model generating in 1-4 steps.
That is not "turbo" model whatever yo say.
@mirek190 Fair enough, you are entitled to your opinions.
TBH, after reading your opinions and looking at your profile, I couldn't care less what you think either 😂.
I made some grids of all currently uploaded turbo models. all but the base model and jib's need atleast 3 steps for a decent image, and anything over config 2 tends to be either very cartoony or blurry/other artifacts. also almost no difference between 1 step and 2 steps.
(step check 1-4, cfg1)https://civitai.com/posts/929205
(cfg check, 4steps)https://civitai.com/posts/929178
(same model order as pervious grids, cfg 1 steps 3) https://civitai.com/posts/929237
This is an excellent model for quickly testing out prompt. I use DPM++ SDE Karras/CFG: 2/5 steps.
This allows me to cut my generation time dramatically without losing much quality.
only 247 downloads of one of the most capable checkpoints in all of SDXL for v1.0 is an absolute crime.
Honestly i ran another x/y/z plot comparing checkpoints again today and Phoenix (original v1.0) is just so ridiculously underrated. The outputs are phenomenal. Still only 266 downloads as of writing this, and it's absolutely a top-5 photorealism-capable NSFW checkpoint among all of SDXL. An absolute travesty that it doesn't get the praise it deserves.
Looks like we don't have an active mirror for this file right now.
CivArchive is a community-maintained index — we catalog mirrors that volunteers upload to HuggingFace, torrents, and other public hosts. Looks like no one has uploaded a copy of this file yet.
Some files do get recovered over time through contributions. If you're looking for this one, feel free to ask in Discord, or help preserve it if you have a copy.