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
TurboVisionXL Version 2.0 Release Notes
I've developed a better pipeline for updating this model while preserving its ability to generate good images at usable resolutions and aspect rations. I've got it pretty solid at 1024 x 1024 now, though you still have an increased amount of twinning and weirdness when pushing towards the 16:9 and 21:9 ARs. 2:3 and 4:3 work pretty well.
Please note the style of this model is still a bit in flux. While I mixed it originally off of DynaVision, my goal is to make this model more general purpose and capable of pushing other art styles. The training in this version has added more photographic and cinematic quality, but still produces fun chibis and anime style output with minimal prompting.
This version produces "usable" output as low as 3 steps, but 5 steps is still the sweet spot. Regarding samplers, I've found you can use the DPM++ SDE sampler in Auto1111 and it produces similar quality output to DPM++ SDE Karras but faster.
ChangeLog 12/4/23
Merged with an (internal only) turbo-ized version of NightVision which brought in improved coherence and better handling of wider AR's. also seems to have mostly stabilized 1024x1024 output.
3 new trainings on cinematic photography, widescreen AR images, and more coherence focused tuning.
finished with a light backmerge with a tweaked version of XL Turbo.
Known Issues
coherence is still wobbly, but improving rapidly version to version. Almost as coherent as my mainline models now.
twinning on non-square ARs can occur
hands, male genitalia and distant faces
Doesn't work with refiners (just not enough steps to give a refiner time to work)
issues with ADetailer...? (reported from comments, I can't confirm this - use Face Editor instead if its broken for you, Face Editor works fine with Turbo in my experience)
Multipass (3M and 2M) samplers are hit or miss - they're just not able to resolve fast enough versus the other samplers or they introduce errors on the multipass. best stick to the single-pass samplers for Turbo.
Restrictive Licensing - SDXL Turbo is being held on a research license by SAI currently. This is out of my hands - I choose to honor open source licensing rules, and thus, I cannot make this model available for onsite generation, as that is against the license on the base model. As soon as SAI removes the license restriction, I will allow monetization and generation options!
FAQ
Comments (13)
the best i try
TurboVisionXL Version 2.0 Release Notes
I've developed a better pipeline for updating this model while preserving its ability to generate good images at usable resolutions and aspect rations. I've got it pretty solid at 1024 x 1024 now, though you still have an increased amount of twinning and weirdness when pushing towards the 16:9 and 21:9 ARs. 2:3 and 4:3 work pretty well.
Please note the style of this model is still a bit in flux. While I mixed it originally off of DynaVision, my goal is to make this model more general purpose and capable of pushing other art styles. The training in this version has added more photographic and cinematic quality, but still produces fun chibis and anime style output with minimal prompting.
This version produces "usable" output as low as 3 steps, but 5 steps is still the sweet spot. Regarding samplers, I've found you can use the DPM++ SDE sampler in Auto1111 and it produces similar quality output to DPM++ SDE Karras but faster.
ChangeLog 12/4/23
* Merged with an (internal only) turbo-ized version of NightVision which brought in improved coherence and better handling of wider AR's. also seems to have mostly stabilized 1024x1024 output.
* 3 new trainings on cinematic photography, widescreen AR images, and more coherence focused tuning.
* finished with a light backmerge with a tweaked version of base XL Turbo.
Known Issues
* coherence is still wobbly, but improving rapidly version to version. Almost as coherent as my mainline models now.
* twinning on non-square ARs can occur
* hands, male genitalia and distant faces
* Doesn't work with refiners (just not enough steps to give a refiner time to work) - I may be able to create a turbo refiner, but it's very low on my priority list right now.
* issues with ADetailer...? (reported from comments, I can't confirm this - use Face Editor instead if its broken for you, Face Editor works fine with Turbo in my experience)
* Multipass (3M and 2M) samplers are hit or miss - they're just not able to resolve fast enough versus the other samplers or they introduce errors on the multipass. best stick to the single-pass samplers for Turbo.
* Restrictive Licensing - SDXL Turbo is being held on a research license by SAI currently. This is out of my hands - I choose to honor open source licensing rules, and thus, I cannot make this model available for onsite generation, as that is against the license on the base model. As soon as SAI removes the license restriction, I will allow monetization and generation options!
This is incredible, I'm really surprised with the result with just 3 steps, it works with controlnet with canny, image to image, and my own custom-trained lora from SDXL 1.0, which works with this turbo model. Thanks for making this, keep up the amazing work!
You should change the name to "turbo merge" because it is not a real "turbo" model.
Cheers
Results at different CFG, Steps & Samplers
Here's a more photoreal prompt showing the same. (TVXL V2.0 BakedVAE, RTX3090, 1024x1024, 7sec./image, Auto1111 )
I had to google YMMV. I hate it, can't you ppl use words ffs lol
Hello author, I couldn't find any training code for sdxl-turbo in the generative models repository of Stabilityai. May I ask what method you used for training, not the ADD method proposed by Stabilityai?
I could use some help, please. I'm using this checkpoint on Automatic1111. When creating images, I can see in the preview that the images are trending towards a nice looking image. But then as soon as it gets to the VAE decoding... BLECH. The images turn into garbled dots of RGB. Any thoughts on what my problem could be? I've selected "None" for SDVAE. Cfg is 2.0. Steps are 5.
I'd like to try this on tensor dot art, but there is some kind of error there: The format of the model file is wrong, needs to be re-uploaded.
Honestly, I'm amazed at the quality of this model. Compared to other models, even with big names, the results with this one are simply incredible (and ultra fast). Thank you so much for your hard work! 👌
By far my favorite SDXL model now. Versatile styles, stable output, better quality than most other models, works well with LORAs and ControlNet, no need for extensive negatives or quality words in the positive, and all in only 5 steps at <10 second generation time on a RTX2070, amazing.
this one is amazing. I tried a few other popular turbo models but this one really works and looks amazing
What settings did you use to train it? And did you use kohya/scripts? I would like to try some ideas and collaborate if experiments succeed.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.