This is a merge between the XL 1.0 base model and the refiner, also I merged in the offset Lora directly. In my opinion it can sometimes create better images than the normal 1.0 base model. Also it works better with fp16. All images here are without the refiner.
The original can be downloaded here. https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0
This is my first SD 1.0 Merge. The next one will be a new trained model. Give me some feedback if you like this one.
Check out my other projects if you want ;-)
Thanks to freek22 who helped my creating this. I highly recommend his new checkpoint Fenris.
Update: I managed to create a new merged checkpoint based on XL3. It's called XL6. I think it performs better than XL 3. But it will give it a certain style.
XL3 is a very good base to create a new model. Trained or merged.
Description
First version
FAQ
Comments (31)
Looks promising!
Thanks. this needs more testing though :-) I haven't tested it much properly. Send me some results if you like. Maybe I use this checkpoint as base to train my next checkpoint.
thank you!
i dont know wich kind of woodoo alien tecnology do this but.. did you notices that refiner get rid of those small dots that comes out on clothes and randomly here and there? like on rocks in the image you posted below
well yes this is awesome.
i think those dots are related to the water mark
not against your work, but SDXL is a piece of shit.
well right now my SD 1.5 models are better than SD XL :-) I see it this way: the first SD 1.5 looked like crap too. This is more like a clean canvas to me. And yes its sometimes poop right now..
SDXL is miles and miles beyond what the base models SD 1.4-1.5-2.0 were. Therefore, it will be a lot easier to train/finetune on top of, which is what this community is so enormously creative at doing. In a week or a few months at most, people will have trained models on top of SDXL which will blow your mind!
Just go back and try do to something decent involving female anatomy on base SD 1.5, and then tell me again that SDXL is a piece of shit :D
People like you should be banned from access of all SD-XL models and community.
@seedstyler Thumbs up. I for one appreciate your constructive feedback and your numerous contributions to the Stable Diffusion community. It seems nobody is brave as you are to tell the truth.
@BitShifter Right now I tend to believe him. I had a hard time to actually work with it. Normally I use SD.Next but for me it doesn't work at the moment. Then I have random crashes. I needs much more VRAM right now than SD 1.5. I did a comparison between XL3 and Colossus 6.11 and Fenris. I mean the results of XL3 were sometimes really good but overall Colossus won the battle. But here is the other side... I am getting my SD 1.5 checkpoints hardly better. I see lots of potential in SD 1.0XL. The software will getting better soon. There will be much of inprovement regarding performance. Also I can't wait for new checkpoints the community is creating right now.
If you use the same prompts as you did for 1.5 and do not set the sliders properly, the problem is you, not the model.
All of you are totally misunderstanding the OP. All he is saying is that he appreciates all the time and effort that content creators invest in order to improve our output beyond the base models. Like how the output for SD 1.5 was just so/so and required a lot of VRAM initially. Think of this more like a pep talk or tough love.
@Afroman4peace The random crashes are SD.Next's problem, not SDXL.
For me SD 1.5 has been already cranked up to the limit. It can't improve anymore. It's very good but it's not on par with Midjourney. SDXL has the potential to be on par with Midjourney. Just watch at the colours....SD 1.5 can't match that fullness, vibrance and intensity even with noise offset.But for me the work will be lot more because there will be lot less people in improving and mixing it:without optimizations the VRAM requirements are too high for more than half of the people here.SDXL could probably fail as much as SD 2.1.768 because without a devoted big comunity it will not be improved neither adopted even.Nsfw censorship is the usual big problem: it's crippling even dressed human silhouette reproduction.Then in this phase every single SDXL image seems lacking both definition and crispness:they all seem 2005-2009 videogames snapshots ...when the "bloom" effect was the thing...
@sambaspo you are absolutely right here. I think my SD 1.5 models reached their limits. Loki 2.0 and Colossus 6.11 represents for me my personal maximum. I also use them parallel or together with SD 1.0XL. I will later post some pictures to show how powerful this combination is. The only thing that holds SD 1.0XL back are the hardware requirements/ or the lesser optimization. I have a 4070ti and I am struggling with getting pictures over 1200x1200. As I mentioned I tested my best SD 1.5 models against my new SD 1.0 XL. In graphics alone my 1.5 models are better than 1.0xl without a doubt (at least in some parts) Reflections are better with 1.0XL. But and here is the important part. In variation, design , coherence or just following the prompts XL3 and XL6 are far superior than any of my SD models.
Man that was a stupid comment. The base model needs to be trained that's all.
@BitShifter The comment more looks like "I just can't get any good results with SD XL so it must be bad." to me.
I guess you might be the ones that just use it for NSFW and women. Because for generating Art, SDXL is amazing with all artists styles and composition, creativity, for me beats all SD1.5 models. I finally can generate creative art, feels much closer to MJ now. I'm just waiting for a finetune to miraculously resolve the deformed hands :)
I tested this against various SDXL models:
SDXL Base, SDXL Base Vae fix, MBB, Dreamshaper, TdG8uU.
This one is the overall winner!
Thanks for testing. this is quite a suprise. I just tested it against the base model and my own SD 1.5 checkpoints
Can you post a comparison between them? This would be very interesting.
@Afroman4peace I posted an image. Overall, XL3 seems the best at composition and hands. Dreamshaper is good too.
This is very interesting. Thanks for sharing this with us. So it seems to be that XL3 don't need a Offset Lora anymore ;-).. And additionally it seems that merging in the refiner did also had some effect.. (it changed the picture when compared with the Base+ Lora..
@Afroman4peace Maybe you could get even better results by using the "SDXL Base Vae Fix" and the "SDXL Refiner Vae Fix". Those seem to be a bit better than the standard versions.
@Daedalus_7 I have to test this more tomorrow. But now I need to sleep. I worked through the last night yesterday because I was so intrigued by my experiment. To be honest I don't know exactly how it works but I just tested it out until I was satisfied with the results.
@Afroman4peace you know the same for me I was surprised that it worked that good and even my sd1.5/sdxl hybrid seems to work so far
@Daedalus_7 I just released a even better checkpoint. It's called XL6.
@Afroman4peace I am using it now :)
FP16 means it is easier on lower VRAM GPUs?
basically yes. Its also easier to train. To use this right you need to download this VAE https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main Also you can add --medvram to reduce VRAM use