Like the work I do and want to say thanks? Buy me a coffee or Support me on Patreon for exclusive early access to my models and more!
Wow guys, you have to try this for yourself! 😲😲
Super fast generations at "normal" XL resolutions with much better quality than base SDXL Turbo!
Suggested settings for best output
Sampler: DPM++ SDE or DPM++ SDE Karras
Steps: 3 - 5
CFG: 1 - 2.25
You can run this model in Automatic1111 like a normal XL model, however not all samplers work with it. I've found DPM++ SDE is the best output performance in the 3 - 5 range, while DPM2 looks really good in the 6 - 10 step range if you're willing to wait just a bit longer. The LCM sampler and Euler A produce almost identical output, which is usable at low steps, but really lacking in detail vs. the other options.
LoRAs work fine, I've tested multiple LoRAs and they appear to still produce expected results, tho YMMV of course.
Coherence is a work in progress with this model. 1024 x 1024 is pretty well solved now, and I very rarely see errors even in 4:3 and 2:3 formats. 16:9 does result in some twinning, but it's not too bad. 21:9 is rougher with an annoying amount of twinning and errors, though not too much worse than normal mainline models.
NOTE ON LICENSING - This model is based on the SDXL Turbo model released by Stability AI. They have flagged the model as being released under a non-commercial research license and permits personal, non-commercial use only. Be aware this model cannot be used for image generation services at this time. If you have questions, please reach out to me on Discord.
Description
Alpha Version 0.1.0.1 Release Notes 20231128
This model is all about SPEED - It produces better quality output than using the LCM or distilled models, works at 1024 x 1024 (and other "standard" XL aspect ratios) and produces high quality output fast. This model is based on DynaVision and will produce stylized output. I will be working on another version that will target more photographic realistic style output.
Known Issues
Coherence has taken a hit - fingers, hands, extra arms. This stuff will show up now, it's something I will work on fixing with updates
doesn't support all samplers, stick to DPM++ SDE Karras for best results!
requires really low CFG! 1 - 2.25 is best as tested
FAQ
Comments (20)
is this turbo? or what does it have on top
hmm, so far running at normal speed, but sometimes new models take a little while of using to 'warm up,' once installed. One weird thing - most of the images have a normal amount of arms but random extra legs. Guess I'll see if adjusting my neg prompt can fix that. Seems cool in theory!
I'm still tRYING TO WRAP MY HEAD AROUND LCM and Distilled AFTER i wrap my head around normal XL. XD However, this does look amazeballlssss
can u upload without baked vae?
will images appear before I hit generate?
Holy cr*p, Batman! Aside from the photorealism, this is freaking awesome!
Can you please enable the model to allow it to be used for civitai's online image generation? Thank you for your consideration.
NOTE ON LICENSING - This model is based on the SDXL Turbo model released by Stability AI. They have flagged the model as being released under a non-commercial research license and permits personal, non-commercial use only. Be aware this model cannot be used for image generation services at this time. If you have questions, please reach out to me on Discord.
Could you share your workflow? For some reason I am getting really bad results even after setting up the same suggested settings.
Please add it for onsite generation 🔪
YOU MUST USE 5 OR LESS STEPS
YOU MUST USE 3 OR LESS CFG
Normal SDXL LoRAs should work and may actually help consistency.
In the images section I posted a screenshot of my A1111 settings page & sample images created in 100image batches using a couple of LoRAs. With these new Turbo models you must use a totally different range of settings - 5steps / 1,2,3 CFG - and a LoRA is a good place to start.
Used a 4090 and could finish a batch of 100 images (1344x784) in a bit over 3minutes. I could probably generate 100images using a normal model w/ 20steps in 6ish. So it's up to you to decide if that speed difference makes the "different" quality worth it to you (Obv may make more or less of a difference on less powerful GPU).
Good luck!
AI development speed is terrifying. Because of low VRAM, once I thought SDXL to be unusable on A1111, but now I can.
Performs worse than standard sdxlTurbo for me with same inputs/seed, even trying different schedulers and cfg
Thanks for sharing, I hope to continue to update and get better results!
Nice model! It does have some coherence issues with 1024x sometimes but I'm getting some great quality outputs with 13 steps, 2 scale, euler_a sampler
This is great, nice job! getting good results at 1024x1024 with 8 steps using euler or euler a on an RTX 4080
What's a good VAE to use with this? My results are really washed out in general.
Amazing results. thanks for sharing.
for comparison, I created this flow.
https://gist.github.com/Dav1nx1/d86ba2c3f96298192bca52cf950e16ab
New version coming tomorrow! 1.3.1 Beta, my patrons and testers have been using it and reporting much better coherence and better quality output. Coming soon! 😉
Works great with animation (Deforum YT clip: Sharon Stone/Alice Eve + Chuck Norris vs a beaver): https://youtu.be/AE5R907X_mQ
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.