Merged with Hyper dev 8 steps lora from Bytedance, Unet only, need clip L, t5xxxl and vae
https://huggingface.co/ByteDance/Hyper-SD/tree/main
Ultimate version
More versatility
Improved artistic side
V4.0
Better realism
Better versatility
Better nudity
Try it with:
Euler/Beta, 10 steps, guidance 3,5 and detail deamon
V3.0
More lora merged for more versatility
Better lighting
Fix vertical trame
V2.7
Merged with my lora: https://civarchive.com/models/866492/moreface-lora
More realism
More versatility
V2.6
More realism, thanks to @bissonfrederic69429 for his wonderful lora SkinDetails
Recommended settings : 10 steps, Guidance 3, Euler/Beta
V2.4
More lora merged
Improve NSFW
More realism
More versatility
Hyper Dev gguf-q4_0 version merged with my lora Real-lora.
Hyper Dev Fp8 Unet version merged with my lora Real-lora.
Hyper Dev bnb nf4 Unet version merged with my lora Real-lora.
https://civarchive.com/models/650205?modelVersionId=727461
More realism for skin, background with less blur.
8 to 20 steps recommended settings : 12 steps, distilled CFG 3, Euler/Beta
Node for comfyui: https://github.com/DenkingOfficial/ComfyUI_UNet_bitsandbytes_NF4
Work with Forge, add vae and clips manualy.
Description
More realism for skin, background with less blur, 8 to 20 steps
FAQ
Comments (39)
Wow this looks pretty interesting! Any chance for an FP8 version?
Yes fp8 coming soon
Can you also add a Q4_K_S version ?
It's about the same size as NF4 and it has better support in Comfy UI.
@jr81 i dont know how to make gguf, but i try to learn
@jice Thanks for the fp8 version! Been using it for the past few days and it really does give nice output :)
A small heads up about the gguf. Q4_0 is considered legacy as the newer K quants provide much better quality. See this post on github explaining the different levels:
https://github.com/ggerganov/llama.cpp/discussions/2094#discussioncomment-6351796
Not sure how you created your gguf but if you used the tools provided by city96 there should be an option for the newer formats. My personal preference is for q4_k (it's the same as q4_k_m) and I used the tools from city96 to convert the fp8 you gave for my personal use. If you aren't sure how to convert to the K quants give a shout and I'd be happy to help.
Do yoy have a tutorial in hand on how to do it? 😍
@redsparkie To do what?
Could not load it in Comfy with neither bnb nf4 loader, nor unet only diffusers loader :/ Getting this error :
"ERROR: Could not detect model type"
SOLVED → https://github.com/DenkingOfficial/ComfyUI_UNet_bitsandbytes_NF4
@jice I had the same error so I installed that custom node and switched it to use the UNETLoaderNF4 node. I does load the model but then bombs out in the sampler with this error:
Error occurred when executing KSampler: 4-bit quantization data type None is not implemented.
Maybe not compatible with the regular KSampler and I need to use the SamplerCustomAdvanced?
Yeah, my bad, that was the issue. I didn't look at the github readme close enough because it has a workflow that uses the SamplerCustomAdvanced. Thanks @jice !
Yeah, I was just blind and tried the wrong one in different places thinking it was the other, mea culpa 🤦🏻😅
Working beautifully now; going to recommend your checkpoint on my next workflow update!
@redpinkretro @Grumblebutt Have fun !
cloned it into custom nodes and couldn't find it. turns out i had to read the requirements text file in there. seems like i had to pip install bitsandbytes>=0.43.0 and then load the workflow at the github page. that seemed to do the trick.
I created a ComfyUI workflow for this that also provides upscaling, https://civitai.com/models/700264?modelVersionId=783547.
Nice work !
I verified it works fine with the FP8 version as well, https://civitai.com/images/27077831.
ForgeUI error (AssertionError: You do not have CLIP state dict!)
@mrkhusbu You have to add the clips manually in VAE / Text Encoder dropdown menu of forge
@jice Please let me know which clip to add
@mrkhusbu look the comment hoodady has posted the link for clips and vae
@jice yes, workng.... Thanks!
Fantastic, can you make a merge with the NSFW too so no need to use lora for it ?
May be if i found a good work...
To use the unet, select/add the following encoders in the Text Encoder Setting of Forge.
Put clips and t5 in text-encoder folder of forge and ae to vae folder
Do Loras, in general, work with this and the other checkpoints you made?
@jice Thank you, Do I have to do anything special? I tried lux Dev person Lorsas and they didn't do more than very vaguely influence the faces. On a side note, your Flux models have great skin tones.
@BiffLohman I don't know this lora, maybe try different weights?
It just worked, this time. I guess things can be a little random sometimes.
@jice It's really disgusting that loras are not supported for nf4 in comfyui. This model is amazing and very fast!
@astragartist a fp8 version coming soon
@jice great news! Thanks in advance!
I see. I noticed in the png info of one of your samples that you used bnb-nf4 (fp16 LoRA). Now, the Loras worked perfectly. Nice!
@astragartist she is there
This is an amazing model! It's really very fast and with real lora results are breathtaking!
Thx for your feedback
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.