💖 Support My Work & Explore More
If this model is useful to you, your support helps me keep publishing free models and open-source AI tools.
Patreon: Support on Patreon
Ko-fi: Buy me a Ko-fi
Buy Me a Coffee: Buy me a Coffee
Discover more of my models, projects, and practical AI resources:
Portfolio: artificialguy.com — models, projects, and experiments.
Prompt library: FindGoodPrompt — curated prompts for image and language models.
AI tools directory: JustGoodAI — a curated directory of useful AI tools.
Twitter/X: @artificialguybr — new models announced here first.
Pitaya.Redmond is here!
I'm grateful for the GPU time from Redmond.AI that allowed me to finish this model!
This is a generalist model fine-tuned on SD 0.9.
The model has a high capacity to generate realistic, artistic images, cars, people, and a wide variety of themes. It's a versatile model. Its a little improvement from already good Stable-Diffusion XL 0.9 Model.I am not 100% satisfied with the model at the moment.
This model is being published as a concept. It was conducted as tests to validate the settings in the weeks we had access to the researcher weights.
The model does not have a big difference to SD XL 0.9, but improves some results. The high quality of the base model and its architecture ends up making the model more difficult to be much better.
This model has been trained for approximately 700,000 steps. In the training process it was realized that the XL model needs a high amount of training to have big changes. EDIT: Joe Penna commented SD 1.0 will be easier to train.
I really hope you like the model and use it.
This model follows by determination of Stability.AI the license of SD 0.9.
The model was authorized and its release as a concept. Authorized by Joe Penna!
Description
FAQ
Comments (12)
Doesn't this need a config file?
No.
omg.. 700 000 steps for minor changes and a 13 gb file.. I can already see sdxl not doing well on here at all. plus it does seem that it wont do nudity?? ive already seen it screw up body proportions as a direct result of that
1 - The base model is much better, it's hard to do something that people see as a significant improvement. I've had models that changed more, but they didn't win in my a/b blind tests i did with my friends and others users.
2 - The model was trained with settings we know today. The SD 1.5 models didn't start out being very good, in fact they were pretty bad. I believe we will discover better ways/settings to fine-tune the model as time goes by.
3 - I don't see the need to do a lot of fine-tuning on the base model that is already good. Lora works and works very well, it's much more worth the effort to train LORAs than to train fine-tuning considering the necessary machine + the quality that is coming out in the LORAs.
Can this model do nudity? Curious.... Training data affects this? I am a noob here......
its one of those catch 22 situations.. it might be better but at what cost? it take way way longer to train, you need way more beefier hardware to run it and i suspect it will take longer to generate images. files will be bigger, taking up more space etc etc. its a cascade of problems which translates to the following. if your average guy can hardly keep up with current tech and hardware, how will they keep up and support or develop the new tech when they might have to invest a whole lot more time and money into this. pity these models cant be updated with current trained models but will need to redo everything from scratch. i might be wrong with my thoughts and concerns so well see..
Thank you for your work here @artificialguybr! Your experiences (and a few other finetuners testing out the model) is one of the major changes between v0.9 and v1.0
The 0.9 research weights were difficult to train.
We're sure that 1.0 is going to be a lot easier, since we spend a lot of effort to remove tons of local minima that the model had found.
To the point of extreme distortion, even. Check out the prompt joe biden on SDXL 0.9 base model, and you'll see why 700K steps didn't do much for you.
That, plus the fact that we're seeing learning rates small datasets are vastly different than SD 1.5.or 2.1
Thanks for all the support and the opportunity to test the research model! Looking forward to do even more and better in SD 1.0.
Keep up the excellent work. Congratulations to you and the whole amazing team.
Very nice Model. I am curious to see what the future will bring. SD 1.5 and 2.1 I have deleted.
Very interesting
But this is not a "base model" it's a merge
The size of these models is a bit disconcerting. I can't even try this right now cause of bandwidth limitations, it'd take all day and I got stuff to do.
But I do want to add, that there seems to be very clear benefit over the base SDXL prompts used in the comparisons in the cinematic realism area.
I love your generalist models. I did some amazing realistic models training it further and merging it with some good versatile models meant for people. I'm not going to try this one though, I'll wait for the SDXL 1.0