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 V3.2 Public Release Notes
I continue to work on improving coherence across wide aspect ratios, getting limbs and hands and faces under control and improving overall image quality at low shot (3 - 5 step) values. For this update I worked in several of my art tunings along with working in some of the latest NightVision training. The model is really starting to develop its own style and feel now.
Change Log 12/8/23
NightVision 7.9.1 tunings added for better photographic output
addition of several of my internal art tunings that help bring out some extra detail and creativeness for regular generations
Known Issues
Widescreen coherence (16:9, 21:9) continue to be a weak spot for this model. Expect coherence errors and twinning on widescreen output.
fingers and hands are still worse than my mainline models, but improving version over version
faces at a distance can be mushy
FAQ
Comments (9)
A quick note on why I label TurboVisionXL as a "Merge Model"
I have labeled this model a merge, because it started with exactly that, a simple merge between DynaVision and SDXL Turbo. Since then I have applied roughly 20 or so tunings from other trainings on my other models, and am now doing model specific trainings to target and fix shortcomings (like bad widescreen coherence). to be clear, I am tuning this model just like any of my other models with my own inhouse trainings or with trainings from my discord community. This is not just "some merge" of my other models, it is it's own unique thing.
I target my tuning and training at 3 - 5 steps on this model. As you can see in these x/y's, quality really peaks around 6 steps (but is still entirely useable at 4 and 5 or even 3 steps if you're just proofing). You can continue to add steps of course, but my focus has been and will continue to be trying to push this model generating high quality faster and faster. I have another exciting update in early testing now that corrects much of the widescreen coherence issues, watch for that in the next week or so. ❤️
Oh, and those few who keep complaining that this is "NOT A TURBO MODEL" and that it's mislabeled because it doesn't generate in a single step... Due to the restrictive SAI licensing of the SDXL Turbo model this is built on top of, I must keep this model in the SDXL Turbo category. It's a licensing thing, sadly.
Once Civit has set up their turbo licensing deal with SAI, I will be more than happy to enable onsite generation.
Is it possible to use this with AnimateDiff in auto1111? I get an error about not being able to use XL with AnimateDiff
For those who haven't thoroughly tested it yet. LMS Karras seems to be the way to go. It's the best combination of speed and quality with such small step count. (at least on my garbage rig). Not only did the quality beat out many of the other samplers, it was the fastest as well.
tested with all available samplers in auto11 with step count of 5. I haven't scaled the slower samplers with fewer steps to compare quality on a "like for like" basis time-wise. there were a few that came out with incredible results but ended up being slower than just using normal SDXL with a fast sampler.
thanks for this amazing model, it really makes SDXL feasible for such hardware challenged users such as myself.
such a cool model
This checkpoint is absolutely amazing.
Does BREAK work for these turbo and xl models? It doesn't seem to do anything but maybe it's a comfyui thing I dont understand
This model doesn't seem to work very well when refining planar plans in Graph-born graph mode
The color temperature is a little low, and the skin color is generally too yellow.
I think the color of ProtoVision v0.6.2 is the best.
This model is the best one among the SDXL models I’ve tried. It renders fast and works really well even with low step values. Despite having a low CFG scale, it doesn’t produce strange images. Instead, it reflects the prompts I entered well and generates images. And the image quality is good. I want to compliment the person who deployed this model.
Details
Files
turbovisionxlSuperFastXLBasedOnNew_tvxlV32Bakedvae.safetensors
Mirrors
turbovisionxlSuperFastXLBasedOnNew_tvxlV32Bakedvae.safetensors
turbovisionxlSuperFastXLBasedOnNew_tvxlV32Bakedvae.safetensors
turbovisionxlSuperFastXLBasedOnNew_tvxIV32Bakedvae.fp16.safetensors
turbovisionxlSuperFastXLBasedOnNew_tvxlV32Bakedvae.safetensors
turbovisionxlSuperFastXLBasedOnNew_tvxlV32Bakedvae.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.