While this model is still available I highly recommend using this one instead
Generate 1k images and natively upscale up to 2k with this lora trained with BF16 and 2x 3x the latent size of standard SD model.
The examples are from a custom merge and juggernaut.
All 2k images are text to image generations that where up-scaled via img to img set any where from .3-.5 using Karras or even the standard sampler.
A 768x1024 image of the face text to image example was also uploaded.
I was not aware of a similar lora to this, however it would make sense as many people who are training XL would probably have the idea to just use that set of images and see what would happen on SD
For this Lora the math (U-net and clip) is geared toward generating nude but not pornographic images of a female.
I posted an example of a CPU only generated image on Auto1111 done with a older CPU at around 5 seconds per IT. Average GPU can do a batch of 4 in 20 seconds or so, but this was a proof of concept for CPU generation.
(Even with CPU's not handling BF16 they can process FP16 and FP32 decently)
Description
FAQ
Comments (6)
Hello, thank you for your contribution however I cannot understand how to use this. Can you give more detail abhout the exact workflow and steps to achieve your results? Thank you!
A lora value of .8-1.0 can be used for generation of 1024x1024 or 768x1024, with the same lora value up-scaling by 2x can be done at around .3-.5 (Karras is recommended) any of the .png files can be imported into comfy to see generation info. I would recommend the face images as they have both the initial and up-scaling done in juggernaut
@Felldude Thank you, is it also used in A1111? IMG2IMG, add LORA: 0.8 as a positive prompt word, and then enlarge the image, right
@wyxzddsjj919 I have not tested in A1111, but I would use the prompt in the positive in both steps at a fairly high value yes. I might do some A1111 testing tonight to verify
So this is similar to the HD helper model by Ostris but specifically for nude woman?
**I have tested it, it does show strong improvement on generating 768*1024 images.
It appears yes, although I wasn't aware anyone had tried it before, I was preparing a XL training and thought I would see what would happen if I trained SD.
But yes because a lora needs to be limited in scope of what it trains unlike a checkpoint this one will be geared toward a nude female model, and thus is in the adult category even though it is non sexual
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.