Goddess Project Lite (1.5GB)
Distilled model is the first model using distillation of HiDream CLIP-L
As the models grow grand, so vast and so wide,
My thoughts turn to those who have little to bide.
Not all can enjoy FLUX or the power of Mochi,
Nor SDXL or PONY, so sleek and unholy.
Tis the season of joy, with its festive cheer,
I think of those who may not have it near.
What better gift, to spread holiday cheer,
Than huge tits and vats of good frosty beer!
Highly tuned model with FP32 TE and FP8 UNET
Finetuned mostly on beautiful females, epic landscapes with some NSFW capability.
768x1024 or even 816x1232 capable
Easily run on a 2GB GTX or even a 15 year old CPU only build
All images use the 1.5GB model
Description
FAQ
Comments (26)
Even if I started directly with SDXL, I think it's good to make this type of support.
I think I'll try SD1.5 just to get the girl with bikini and puppy from the samples.
942733261732203, "steps": 25, "cfg": 4.0, "sampler_name": "euler", "scheduler": "exponential" - The example image was upscaled with img to img but that is the source image settings
This looks really promising. Thanks for sharing.
Would it be possible to get a version of this that'll work with a1111? It won't load for some reason, I assume it's for comfyui or something...
auto1111 should have native FP8 support, forge also
@Felldude Not on Linux.
Works for me, on windows.
@Felldude I'll try again. Thx again.
How can I not download a 1.5GB SD1.5 model ;D
I always loved 1.5 for its speed and simplicity, and with all the support it got it also makes good images... Hard to believe something so small can be as good as it is.
Dang, it generates pretty girls with just the negative prompt xD
The complexity is there it is just hard to activate without a T5 interpreting natural langue back to the clip-L - sometimes just adding random words helps when trying to get a different image
@punkbuzter340 Yeah if you heat mapped the checkpoint it would be heavily biased toward female with that term being finetuned 100k or more times
Request: Could you add Noise Offset for those more accurate dark images? Night time tends to disappear without it :P
0.03 was used, it could be lack of greyscale and night images in the finetune
I tried using it with Easy Diffusion 3.0.9 and received the following error: InvalidHeaderDeserialization
It requires the use of PyTorch's autocast, as well as FP8 e4m3 support - You could request those features or switch to something like forge or comfy
thank you. for us SD1.5 plebs with 8gb vram.
long live 1.5!!!
Skin is great, the samples seem to be mostly ing2img and when tested as text2img they did not (so far - I have not arcived it yet) stack up against some of the other big 1.5 (realistic, epic etc) hitters (maybe different samplers?)
However side by side testing the skin was way better every time, even with LoRA added it did not come out as well as this model. I am thinking a inpainy of this would or could be great to use with some SDXL or FLUX gens and then the skin from this.
Have not messed with 1,5 for ages but I think we can do an on the fly model merge with inpaint so maybe I can test it that way
No additional lora's where used, it could be the fact the image to image was done natively and not using highrez fix which breaks the image down
@Felldude Actually sorry - the LoRA comment was about other models I tested against this model. They did not yield a skin as realistic but on text2image tended not to mangle body, fingers pose etc. so I added a LoRA to to other models to see if they would come close then I would write this one off as fun to test but no use. However skin was still not as good until the LoRA was high and causing other losses (again not used with your model).
Then when comparing img2img again no LoRA needed with your model, great skin and it seems to excel here in img2img so I feel this could be a great skin replacer.
Sorry for any confusion about the LoRA in my testing
I like the sd1.5 version, thanks. Can i ask if there will be a pony version?
I have honesty never tried the 1.5 pony - I do have version of XL pony that used most of the datasets that finetuned this model
Sorry i mean SDXL Pony version or even just a SDXL version.
@kellykelly The ultrareal 8GB/12GB model uses most of the same training dataset
@Felldude ....I have always been a1111 v1.7 user and sd1.5 user. It was this goddess project that got me to install forge because FP8 didn't work on my a1111 and didn't want to update it. Turns out forge is not bad and so i am switching in between the two. I will download the ultra 8gb and try it, only reason i asked was I was just going by the name goddess project because its trained. I like your stuff.
@kellykelly Thank you, I could have named the pony model that, but my models don’t have quite the popularity for naming conventions….maybe someday lol
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.