SDXL-Turbo is a distilled version of SDXL 1.0, trained for real-time synthesis. SDXL-Turbo is based on a novel training method called Adversarial Diffusion Distillation (ADD) (see the technical report), which allows sampling large-scale foundational image diffusion models in 1 to 4 steps at high image quality. This approach uses score distillation to leverage large-scale off-the-shelf image diffusion models as a teacher signal and combines this with an adversarial loss to ensure high image fidelity even in the low-step regime of one or two sampling steps.
Developed by: Stability AI
Description
FAQ
Comments (36)
Imagine if we can get a SD turbo from the Segmind reduced model... 2024's gonna be crazy for open source AI
except the GPU prices....
@Torquemaster Who needs a GPU? With Turbo CPU based inference is practical.
@Torquemaster Segmind's model is less than 3 gb as whole checkpoint file in fp16, it runs very fast even on my 4 gb of VRAM, if you get the best of both worlds you might generate 1024x1024 images on the worst GPUs even in a few seconds
sd turbo is a new distillation method the segmind models are already distilled if you were to combine these methods more that likely the results would be subpar as you would be compressing an already compressed model
Kinda interesting to see you can get 100 batch of "decent" images in 16 seconds
(～￣▽￣)～
it's only the start. the models are still very poorly formatted.
it's mindblowing
i can't manage to get this model to work properly, i always get a very blurried image with 4 steps only interference, can you please tell me the recommended settings for A1111?
set cfg scale 1 or 2 not higher
use less than 4 steps...
Use ComfyUI + lcm sampler
fyi this is working just fine in A1111 now
Started playing with this model today. The speed is amazing. A few tips I found:
Make sure you generate at 512x512. I was unable to get any other dimensions or aspect ratios to yield a non-mangled image. I actually thought the model was broken at first because the usual1024x1024 generates multiple instances of every subject.
Higher CFG values mostly add contrast. I found anything over 1.5 to be too much. Tiny fractional changes will have a significant impact. A CFG of 1.2 with 2 steps worked best for me on my test image. I got good results with only 2 Hi-Res steps when upscaling.
Negative text embeddings like EasyNegative mostly randomly changed the contrast. I found negative prompts in general to have little impact or to actually degrade the image, although this could vary by prompt. Haven't tried LORAs yet.
Choice of samplers doesn't seem to matter as much as with a standard model. They all produce very similar results unless you happen to choose one that breaks. I liked the images better with VAE set to none. Clip skip seemed to have no effect.
I posted my test image to the gallery. Hope this helps.
So, any SDXL model + lcm module for 1024 is almost the same? I can generate good enough 1024*1024 images in 3-6 steps with lcm lora.
@Shio_N I did a quick comparison between the Turbo model at 512x512 using Hires Fix with Latent upscaling and the LCM LORA with the standard SDXL model at 1024x1024. The LORA needed 4 steps to produce a comparable image compared to 2 steps for Turbo/Hires fix. I found the Turbo model to be about 10% faster than the LORA. This amounts to only a few seconds when run locally with an 8GB GTX-1070 Ti GPU. So probably not a meaningful difference. Neither approach generated an image of high enough quality to be useful to me.
However, using Turbo/Hires fix with UltraSharp upscaling instead of Latent does yield an image that is sufficiently detailed for some situations where quality isn't paramount. It's a few seconds slower than the LORA with this setting.
The LORA is useful if you need a particular model. It also has the advantage of generating more or less the same image layout as the base model, albeit at a much lower level of quality. So it might be useful to quickly hunt for a promising seed. I'm excited to see where the speedier approaches lead as they mature.
Bring turbo to 1.5? Maybe we can start doing more video with faster times.
no
This is a merge of the new SD turbo model with a pre-existing SDXL model there's no reason that the distillation method ADD (Adversarial Diffusion Distillation) which was used to create SDXL turbo couldn't be applied to 1.5 but that would require training a distilled model from scratch using ADD on SD 1.5 currently no such model exists so op can't apply this to 1.5 though depending on the amount of compute required it could be possible for a individual to train one themselves assuming its not to expensive of course.
https://github.com/Stability-AI/generative-models#news
November 30, 2023
Following the launch of SDXL-Turbo, we are releasing SD-Turbo.
@jabberbrillig people will just complain about how it isnt as good
SD Turbo NON-XL Version:
https://civitai.com/models/220609/sd-turbo
Is this the same LCM-LoRA integrated in this SDXL? Or we can use this Turbo model with LCM-LoRA to go even faster?
NSFW not possible, right?
I can't get it to work with NSFW LoRas.
Of course it doesn't work. It is the official version from stabilityai. They don't like NSFW
@TheDarkLurker Let's see how long it takes the community to create some neat XXX-XL-Turbos.
@Bl4cku5_H34du5 check out my newly released turbo boosted version of Ishtar's Gate: https://civitai.com/models/221935?modelVersionId=256868
I made some grids of all currently uploaded turbo models. all but the base model and jib's need atleast 3 steps for a decent image, and anything over config 2 tends to be either very cartoony or blurry/other artifacts. also almost no difference between 1 step and 2 steps.
(step check 1-4, cfg1)https://civitai.com/posts/929205
(cfg check, 4steps)https://civitai.com/posts/929178
(same model order as pervious grids, cfg 1 steps 3) https://civitai.com/posts/929237
It's really a shame that these these links are dead.
my suggested settings for ALL turbo-(merge hybrid etc) models, 2step sample, 2 step high res, cfg:1, Eta:.5 ETA delta:2, always discard last sigma:true, (not advised:extra noise:.25 for more details but lower image clarity), milage will vary
This keeps it at a 4step process and depending on the models "turbo" state should allow for 4x upscaling(2x recommended, .5 denoise hires)
According to your license settings, SDXL Turbo requires credit to the creator when used, is prohibited from creating merged models, and requires the same license when distributing merged models (although it is unclear how they were created). Are you sure ?
Respectfully, I can't get this model to produce anything but burned-out nonsense. Please post some guidelines!
Agreed. It's a complete disaster!
I found this model totally useless. Out of all the other Turbo models this is the only one that not only fails to produce a good image but gives me anime/cartoon style every time. Not worth downloading. I won't give this model a rating when it fails to do what it is intended to do!
Details
Files
sdxlTurbo_fullVersion.safetensors
Mirrors
sdxlTurbo_fullVersion.safetensors
sd_xl_turbo_1.0.safetensors
model_1.safetensors
sd_xl_turbo_1.0.safetensors
sd_xl_turbo_1.0.safetensors
sd_xl_turbo_1.0.safetensors
sd_xl_turbo_1.0.safetensors
sd_xl_turbo_1.0.safetensors
sd_xl_turbo_1.0.safetensors
sd_xl_turbo_1.0.safetensors
sd_xl_turbo_1.0.safetensors
sd_xl_turbo_1.0.safetensors