Hyper-SD is one of the new State-of-the-Art diffusion model acceleration techniques. In this repository, we release the models distilled from SDXL Base 1.0 and Stable-Diffusion v1-5
Project Page: https://hyper-sd.github.io/
Original Repo: https://huggingface.co/ByteDance/Hyper-SD
Original Paper: https://arxiv.org/abs/2404.13686
Description
FAQ
Comments (22)
I didn't have much luck w/ this 4-step LoRA (would love to see some examples posted) - I downloaded their 4/30/2024 8-step from Hugging Face and it works for me wrt getting crisp results.
Just to be clear: the 8-step CFG (Hyper-SDXL-8steps-CFG-lora.safetensors, also here) LoRA worked okay for me.
May I ask if there are any differences in the image quality generated by the 1-step model, 2-step model, and 4-step model respectively?
@dj9wqurf I am not entirely sure but using with ip adapter you get some crazy results ip adapter is super strong in results
Could you please explain the differences and usage of each of these 5 different versions separately?
how do I use this properly? also the cfg model won't show up in civitai helper, does any one have the text trigger?
Thank you for your work, but can I have more information. You realize that KSampler has more than 10 samplers and sampling methods, how to use your model?
I'm trying to generate an image with 2-10 steps gets just .... (it doesn't work very well:)
You have posted the images, please give us information on what and how. Otherwise we can guess until Chinese Easter.
It is not that hard to find out with a xyz plot. It is a bit model dependent, but I use mostly dpm++3m with the uniform scheduler. For some models 2sa or even euler work better. If uniform doesn't work, try karras. I don't use comfy for plots, but the general approach is zo choose your fav model, a fixed seed, and then put the usual modern samplers on the x axis and the schedulers on y. Use standard resolution and the number of steps the lora is named after with a cfg of 1 and no postprocessing. With 2 or 4 steps per image the whole process is pretty fast.
Edit: 12 step CFG version
There is also a 12 step CFG version available. Would be great if that gets added here as well.
How To Use This? I'm Using This With Forge And It Doesn't Work.
So I tried using this in Automatic1111, but I'm not seeing a difference
Do I just drop the lora in the folder & select it, just as any other lora?
Or do I have to do something like installing nodes/workflow stuff?
I cannot for the life of me get the SDXL CFG 8 Step to work. I'm using comfy. Does anyone have a workflow I can glom off of?
I put sdxl 8-step version in lora file folder, but I could not see it at main interface, what's wrong?
How do you use this?
Set Steps: 8, CFG scale: 3, Sampler: Euler a And Use This LoRA By Tag <lora:Hyper-8steps-CFG-xl:1> It will make the work clearer and faster.
can you make it available to use on civitai generation?
can you make it available to use on civitai generation?
Please guys someone make some stubid proof video tutorial how to use this, in forge... i dont find any info about it... only see comfy ui videos...
Use it like any other LORAs.
Set CFG to 1-2 (CFG 1 is the fastest, almost double speed but negative prompt is ignored). set the steps 4-6 for the 4 steps LORA, 8-10 for the 8 steps LORA ... For higher CFG decrease the strength (0.7-0.9)
Hires fix is recommended.
For the 8 steps CFG version you can use normal CFG like 5-6.
I usually use "Euler A" sampler and "SGM Uniform / Uniform / Beta" schedulers.
Drag this image and drop it in Forge into the prompt field, then press the blue button below the Generate button. Change the lora name in the prompt field to match your file. https://civitai.com/images/42682881
what’s the conversion of steps and scale when this is active is normally? i’ve been using a hyper LoRA for awhile but now i can’t seem to find the right workflow. hypothetically if this LoRA was at 1 weight, CFG scale was at 1.1, and steps at 7, what would be the closest to that without this?
You could try LCM sampler with Karras scheduler along with dmd2_sdxl_4step_lora_fp16 added into the mix, it works pretty darn well imo
if u use forge/reforge/a1111, go to settings, in search put Lora, and check "Always show all networks on the Lora page (otherwise, those detected as for incompatible version of Stable Diffusion will be hidden)" and now it will show up in lora.
From my test, Euler Beta is the best, Euler A Beta looks good too, but adds artifacts/noise in some generations so is less stable.
Details
Files
Hyper-SDXL-8steps-lora.safetensors
Mirrors
Hyper-SDXL-8steps-lora.safetensors
Hyper-SDXL-8steps-lora.safetensors
Hyper-SDXL-8steps-lora.safetensors
Hyper-SDXL-8steps-lora.safetensors
Hyper-SDXL-8steps-lora.safetensors
Hyper-SDXL-8steps-lora.safetensors
Hyper-SDXL-8steps-lora.safetensors
pytorch_lora_weights.safetensors
Hyper-SDXL-8steps.safetensors
Hyper-SDXL-8steps-lora.safetensors
Hyper-SDXL-8steps-lora.safetensors
Hyper-SDXL-8steps-lora.safetensors
Hyper-SDXL-8steps-lora.safetensors
hyper-sdxl-8step.safetensors
Hyper-SDXL-8steps-lora.safetensors
Hyper-SDXL-8steps-lora.safetensors
Hyper-SDXL-8steps-lora.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.