DreamShaper XL - Now Turbo!

This "turbo" version is not a turbo version, it's an LCM merge, it can't do 1 step.

i'm already there son

so you've never used turbo ?

Exactly. I just spent the last hour figuring out if any of the others are reasonable, and the one they mentioned is the best. The problem is that there's nothing Turbo about it. It's barely faster than using the regular model with a higher quality and more prompt following setup. The original turbo model was stupid fast with 1 step. All of these others keep getting further and further away from that and are exponentially slower. What's the point?

@Lykon Hey thx for the awesome model. Per image is just 2-3x faster even using the slowest SDE sampler. Was just complaining about the samplers not your model XD

DMP++ 3M SDE provides significantly more details within the same number of steps as compared to DMP++ SDE Karras.

@Lykon what about parameters for Adetailer inpainting?

loading times with xl lora's will depend on your system, enable lora cache if you have the memory and are using the same lora's,

@glitter_fart Thank you, I will check it!

@glitter_fart Where can I enable Lora cache in webui?

use none

It's not trained at all, it's an lcm merge, and not a turbo model. All training data is public information. Anything that needs more than 1step is by default not a turbo model.

needs 4-8 steps, thinks its turbo .......

@Lykon thank you for your suggestions

to be fair, lots of things in this model dont work, like using 1 step

@glitter_fart I see... guess it's better to just wait for it to catch up on SD 1.5

@qazxsw  base turbo is already ahead of 1.5, thats why it can do 1step and this isnt a real turbo model just a lame merge by someone who is very full of themselfs

@Lykon Lmao? Do you normally get these many haters? You are probably one of the better/best custom model builders out there, these ppl must be new.

I hope you enjoyed your vacation btw bud :)

Btw do you know if this will work with FooocusMRE?

Just think where we'll be in a year's time. If they can do this now top work..

@glitter_fart Base turbo looks like shit, just like you.

This isnt a turbo model, it's a merged lcm with turbo, at best. Turbo models, 1step 1cfg, is the base line.

@Lykon Do the site admins just add it for w/e model they feel like?

@Lykon Turbo has a different License than a LCM Version. I am not 100 % sure, but i think it´s cause of the new License from Stability

Can't use turbo models with onsite generation yet,

@Lykon actually you can toggle option for onsite generation. Just ask how someone but it is possible

@NecroBear I understand. Is there date for video onsite generation? 💸

Try 7 steps. I had to play with the balance of the two for mine. I also use the dpmpp_2m sampler.

@Lykon I'm using comfyui and I don't see an option for dpm++ SDE karras. where do I find this sampler?

@Lykon I dont see the samplers in those images in the ksampler node, am i missing a custom node?

@Lykon tnx man :)

@Banebanane a further suggestions: grab the "8m" models from https://huggingface.co/Bingsu/adetailer/tree/main and throw them into models/adetailer (then reload UI). the two 8m models don't come default with the a1111 webui adetailer extension, and I've found them to be quite superior. I suggest raising the detection threshold to around 0.4ish (from default 0.3) for the 8m face model, and perhaps lowering it to 0.25ish for the 8m person model.

Also - when using adetailer for fixing eyes, be sure to turn off CodeFormer or GFPGAN, or it's gonna be bad.

for single subject images or only large-face close-subjects, GFPGAN can sometimes be effective on it's own (i never use CodeFormer, just seems bad all around). But GFPGAN will fail horribly at background faces and completely ruin images, and kind of makes everyone look the same (which may or may not suit your needs).

In my opinion it's better. This is a turbo model so much faster.

DPM++ 2M is fairly deprecated at this point, and it's been found that many common distributions - including the most common DPM++ 2M Karras that comes with Automatic1111 - is literally bugged with an error in the code. There is a couple complex manually-applied community patch options, but regardless, it's probably not worth it. There are equally-fast samplers available with better quality and very similar determinism.

@shapeshifter83 I see ... I didn't knew about this. I only tried A1111 in online demos.

In Comfy UI is what is used the most and never noticed any issues.

But anyways, any other sampler that could be supported would be welcomed. 

I use this to find interesting variations of images I already like. 

I've found some with this model, but they are hard to fix. I mean to get them to the same level of quality of the recommended sampler.

Or perhaps just release a non-turbo version. 

My GPU is on the slower end, but I prefer to wait a few more seconds and get the best possible results.

In the long run, this saves time, versus trying to fix an almost good image that's not quite there yet. 

@jr81 I haven't used ComfyUI yet, but in A1111 there is an option to manually override the scheduler, so the closest alternative to DPM++ 2M Karras would be to manually apply Karras scheduling to the Euler sampler. That would give virtually identical speed and determinism (seeds would look the same) while providing somewhat of a quality boost. The main reason people don't realize DPM++ 2M Karras is flawed is because it's generally the first place people discover the Karras scheduling method which is itself superior and somewhat masked DPM++ 2M shortcomings.

My recommendation with a slower system would be to use Euler (with Karras or even Exponential or Polyexponential, if your ComfyUI has those options) and spam quick batches of 1- or 2-step images with an early approximation of the prompt you want (instead of a full descriptive prompt, just "girl standing" or "man sitting" or "group of people" or "car") and from the resulting shadowy silhouette blurs you can then select a particular seed to try to flesh out into more definition with higher steps and detailed prompt. I find that method faster than running a full prompt looking for the elusive seed that matches the scene setup envisioned in my mind. Generally the samplers will "see" things the same way you do - if you see a face or a man or a car in that vague blur, the sampler will usually see it the same way you do. (it's sort of amazing tbh).

This method won't work as well with "SDE" samplers (because the determinism isn't as stable when changing steps) and won't work at all with "a" samplers. However, to be perfectly honest, I'm usually doing this with DPM++ SDE Karras anyway, simply because I think that's the best overall sampler. Even though "seed shadow hunting" isn't quite as easy with SDE.

Bear in mind certain checkpoints will sometimes require SDE or even DPM++ 3M SDE (Demon Core, for example, and anything merged with that - quite a few around). You'll know it by the resulting strange warping with red tinge.

I hope my rambling is helpful in any way. xD

@shapeshifter83 OK, thanks for the tip.

In Comfy UI, the sampler and the scheduler are selected independently. The DPM++ 2M sampler definitely looks better with Karras scheduler. 

When I have a specific position or overall shape in mind, it's usually from another picture so I tend to use Image 2 image, with different levels of denoising, and sometimes ControlNet. In Comfy UI, this does not increase generation time significantly.

You have to be careful and not ruin your image at the upscale/hires fix stage, I usually go with 0.46 to 0.52 denoise strength. And 3-4 steps. Anymore and you overcook the image. :)

@Dafolie Thanks, I didn't highres fix any of the current experiments. The images I'm getting from the turbo model aren't bad or broken, they're just rather plain/boring compared to what I get from the older model with identical prompt.

I seem to be getting much better results today. I'm not sure what was giving me poor results before. I did have --no-half on, which may have been causing issues, not sure if that was on the other day or not, but it was taking really long time to gen today so I checked settings and after turning that off it's back to seconds to generate and gives decent results. Hires-Fix still takes a long time though, not sure if there are better settings to speed that up, but I seem to have less hand failures with highres fix than without, so experimenting with that currently, despite it returning gens to a couple minutes and image.

I am using a different prompt today than I was when I started this post, I'll have to revisit that prompt to be certain if it is completely better.

Here is an example, that (sort of) shows what I'm talking about:

https://www.patreon.com/posts/which-one-do-you-94628267

Two different sets of prompt, A & B; everything is the same between version 1 and 2, except:
1 = Turbo, Steps 5, CFG 2

2 = Alpha2, Steps 20, CFG 6.5

They're all using the same seed, hires.fix, sampler, upscaler, and size. While nothing is technically wrong with any of the images, the Alpha2 results are much more what was asked for in the prompt, especially in terms of color and mood, and to me are much more creative and asthetically what I wanted and am used to from dreamshaper. The Turbo images are faster and good, but I feel like they're more like what I get from other models, not what brought me to dreamshaper in the first place. So this is what made me wonder if I should be prompting differently, as the colors and style prompts seem to have less effect in the Turbo model.

Well it is because of stability AI, not this dude but I'm sure he appreciates the sentiment! :D <3

@Lykon thanks！So sdxl lcm can be combined to accelerate?

adjust your resolution to be exactly 1 megapixel (1024x1024), first, that will help. Besides that, I don't have any further advice since I don't use turbo, hopefully someone else sees your question and can help.

@shapeshifter83 I see I was running the image a little small but even after adjusting the image to 1024x1024 it still comes out pretty bad. Thank you for trying though, I really appreciate it!

Recommended sampler is DPM++ SDE Karras not DPM++ 3M SDE Karras, if you look at the examples with different samplers all the ones with the sampler you are using are terrible. Steps should be 4-7

oh yea, that's definitely the answer. I missed the 3M in his settings. DPM++ SDE only; Karras scheduling highly recommended and generally always superior. I haven't used Turbo like I said but from what I know about this technology, I imagine Exponential and possibly even Polyexponential scheduling might also produce quality outputs at 4 steps.

@carisrains @shapeshifter83 Thank you both very much!

Hey, do you mind sharing a copy of your google colab notebook? Been having a really hard time finding one that’s working without any problem

@3rixon https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast_stable_diffusion_AUTOMATIC1111.ipynb
and you could use my personal commends which i run everytime with this notebook:
!pip install lmdb !pip install pyfunctional !pip install cchardet !pip install python-dotenv !pip install fake-useragent !pip install ZipUnicode !pip install dynamicprompts[attentiongrabber,magicprompt]~=0.30.4 !pip install torch==2.1.0+cu118 torchvision==0.16.0+cu118 torchaudio==2.1.0 torchtext==0.16.0+cpu torchdata==0.7.0 --index-url https://download.pytorch.org/whl/cu118

I'm using them coz some of extensions doesn't work on colab correctly.

--> 8x_NMKD-Superscale_150000_G
hires steps: 4-5
denoising strength: 2.5-4.5

sounds like an issue with foocus

@fitCorder I don't know, did you test it? Do you have any feedback on the subject?

@Suzanne https://github.com/AUTOMATIC1111/stable-diffusion-webui I get 1 second generations with this model by Lykon using regular webui.

@fitCorder ok, thanks for your answer, but it doesn't answer my question...

@Suzanne Fooocus already have extreme speed mode, I don't know why you need this.

@infrezz721 because I tested the "basic" SDXL turbo and it's even faster, so I thought Dreamshaper would be even better, because the other one has a lot of limitations (square images only, no NSFW)

@Suzanne I mean why you not use Dreamshaper alpha2 on extreme speed mode if this version troublesome

@Suzanne it answers your question that its something with Fooocus

896x1152

Basically every 2:3 (portrait) or 3:2 (landscape) up to 1024x1024 (not 1024x1536, it didn't produce good results consistently).

E.g: 512x512, 512x768, 768x1152, 1152x768.

Yeah, I really want to know a good way to handle this, because I'm also training Loras, and really want it to work with this more than any other model. It would be a shame to throw away all the other models in favor of this though.

Little update: I did something wrong on my lora. After fixing that, it works as good on turbo dreamshaper as on sdxl base. I'm totally amazed about that.

I'm still curious about how this model was trained and what is the best way to finetune it. But I just wanna say that finetuning works in principal, much better than for other models like juggernaut.

Depends on your settings. But most probably it was bad resolution

@Volnovik @Lykon which resolutions do you recommend? Is 896x1152 good with this model? Thanks

@karetirk932  I use following (ready for ar-plus extension):

XL1:1, 1024, 1024

XL4:3, 1152, 896

XL5:4, 1145, 916

XL3:2, 1216, 832

XL16:9, 1344, 768

XL21:9, 1536, 640

Of course you can swap x and y to get 9:16 etc. Also check Fooocus v2 for resolutions used there, there is a bunch of aspect ratios that I don't understand. Also note that aspect ratios above are not strict, for whatever reason Stability AI trained model that way.

No. The StabilityAI EULA restricts using the model itself as part of a commercial product without an agreement from StabilityAI.

For example: You cannot use the model as the backend for an app, wherein you host the model on AWS, the app calls it and generates an image, and you charge people for using the app. In other words, you can't just launch your own Midjourney based on StabilityAI's model.

But you can use it to generate images and you, not StabilityAI, owns those images.

@eoffermann I believe this is the correct interpretation.

It should be added though that "owning those images" is a bit questionable in light that they are not protected by copyright. You are free to use them though, but so are everyone else! (applies to all models, not just Turbo)

@eoffermann thanks for the answer. How about using / hosting the model as an internal tool for production / productivity? So not selling it as a service…

What do we use in comfyui?

I know.

All this weirdness is why I don't really like Turbo models... I don't mind waiting a few more mins for a batch of images to finish vs. Turbo models where like, OK, cool, I saved ~2 mins, but all the images have weird artifacts or don't look as good, so now what? Try to remake them in a non-Turbo model? That wastes even more time than just avoiding Turbo in the first place. RIP Dreamshapers.

Use inpaint mode with high denoise strength.

You might also try with "latent noise" for masked content, low values for "Only masked padding, pixels", and use an ancestral sampler like DPM++ 2s a, instead of DPM++ 2M Karras. Make sure CFG is high enough. Besides, I'd recommend using a 1.5 inpainting model for this task.

Inpainting doesn't really work that well with SDXL checkpoints, even if you used a non-turbo version. I still use the 1.5 inpainting DreamShaper checkpoints for this.

@Lyrinami If inpainting hands or feet, try putting the relevant limb in the negative prompt and leaving the positive empty. It boggles my mind, but it effin works on SDXL.

@steamrick Will try, thx for letting me know!

@Lykon I still love Dreamshaper 8, so thank you for that. It and deliberateV_11 both have almost magical results when I do a eulara pass in img2img.

I personally doesnt have any problems with inpainting on this model.

Works perfect,but only with Foocus,in Automatic1111 not so great,probably because of different inpainting engine.

I try it in many crazy situations,adding all kinds of stuff,improving resolutions and details on faces and even hands,doesnt have any issues whatsoever.

With ComfyUI i dont have many experience so i cant tell.

From my personal perspective model is highly flexible in any situations,from fantasy to realism,i rendered logos,cartoon stuff,comics and SF themed images,and shines in everything.Hands are also almost perfect in any situations so i dont have may cases where i really need to inpaint them for wrong anatomy,i mostly inpaint them for adding details and skin features.

Hope this helps.

@Lykon In general i always turn off all styles in fooocus,even Foocus V2 features because like you said it can damage the images and alter the model.

Im using manual prompting without any of the Fooocus styles,because this is best way to fine tune the model for personal taste and preferences.But for people who are totally new in SD many of the fooocus fetaures can help in great way.

For experienced users i would always suggest to tick off all the boxes and write the prompt in the same way like in automatic1111.

You should either use DPM++ SDE or DPM++ SDE Karras for best results. Besides, make sure you use the SDXL VAE, not the SD 1.5 version. I think the VAE is already baked into this checkpoint, so you might as well choose none.

are you using hires.fix?

Yes, you can run it with the Diffusers library: https://huggingface.co/Lykon/dreamshaper-xl-turbo

@Lykon  I use Fooocus, so it's only the Refiner Switch slider

my question is... how the F someone can tell you "this is made with my checkpoint or with stable diffusion you cant sell" after editing a photo with another software all the metadata is gone... so...

I hope to solve some of the confusion on this subject. They mean companies such as Deviant Art which have their own AI-Generators that use Stable Diffusion. Stability AI etc... the Checkpoint cannot be integrated commercially into their system without the permission of Stability AI.. You however, can use it locally and then your "non-copyright violating images" can be sold by you on other sites etc... such as on DA.... within their rules of course... I hope this cures some of the wonder left by the statement about commercial use. :) P.S. Thank you to Lykon, I have been very pleased with many of the checkpoints you've made!!!

@enochianborg "non-copyright violating images" are ones that don't have some brand's product in them? Or is it more encompassing - such as looking similar to some photographer or artist's work?

@EricRollei21 Well lets say for instance you make a render of "Tom Cruise" that has his likeness... that would violate his "Tom Cruise" rights if you were to "sell" that image without conscent. Looking similar to someones style is commonplace.... pretty much all waifu art is the same style. But there is absolutely NO photography style that is copyright protected. Photography styles have been used and taught in college for years, such that as long as you are not claiming a particular photographer's name or company as if they had endorsed or had a part in it's creation... for instance Pixar or Disney. You could do fan art of Disney's Elsa and post it publicly but NOT sell it for profit. So long that you disclaim that it was "Fan art" and you are not paid for it's exhibition. I have created mature AI generations of a famous persons but I cannot sell them because I do not have permission from the person or their agency which controls their image likeness. Such as Megan Fox, I can post them as "Fan art" but I haven't done this as of yet because I'm not sure It won't get me on her naughty list.... The OP is about the subject of injecting the "TURBO MODEL" into Commercial software that generates A.I. images outside of individual use scenarios to be more separated from what content you can sell as far as your own creations. Like apples and oranges in a way.

What is your resolution ? I get great result for vertical image but 16:9 horizontal is a lot less detailed and sharp.

I'm also seeing quality decline with ControlNet (mostly in the form of blurring)