Goddess Project
--Formerly Uncensored Females--
Standalone Checkpoint - Goddess works in FORGE only
DO NOT LOAD a separate VAE, TE, or CLIP unless using GGUF
This version is a mixed precision Flux Dev model, with limited UNET changes to allow for feminine anatomy.
Run this model in automatic FP16 Lora mode NOT NF4
This model is full precision in BF16 UNET with mixed precision (NF4) on the TE blocks
This model fits in a 24GB card and could be run in GPU only mode as such
High speed BF16 with a slightly lower prompt accuracy compared to the 33GB full model
Links (For GGUF ONLY)
Updated CLIP - Standard CLIP-L - FP8 CLIP-L -- ** Version Comparison **
Per the Apache 2.0 license FLAN is attributed to Google
This model is a training using many individuals with known ages and 2257 forms, it has also been merged to try and ensure that no known individuals can be reproduced. However FLUX seems to like to learn faces even with less then 10% data rather then merge them into a new face.
Description
FAQ
Comments (17)
Hello! I'm new here, how could I use it for inference? Possible to do it on Replicate? Is there a way to train lora on top of it?
Load one of the images into comfy for workflow - You should be able to train a Lora against it, I am not sure how it would train if you used the FLAN T5 as TE number 3 but it would be interesting to see
Thank you! I tried and used this prompt "woman enjoying a cup of coffee. She wears a chic, casual outfit, and her 24-year-old face shows a calm contentment as she looks across the bustling street. The Eiffel Tower is visible in the background, and the golden sunlight catches the details of her face as she sits back, savoring the moment. nsfw" and the output is very nomal :/ wondering if I didn't use the trigger words?
Ah I tried again with "naked" and it worked! Now I have to try to train a Lora against it. I used my previous Lora trained against flux dev but it didn't work :(
I can use the gguf version, but this full version I can't get it to load. It has a different weight architecture than Flux Dev. Wondering why...? Below is the safetensor file, looks very different from Flux
model.diffusion_model.double_blocks.0.img_attn.norm.key_norm.scale [128]
model.diffusion_model.double_blocks.0.img_attn.norm.query_norm.scale [128]
model.diffusion_model.double_blocks.0.img_attn.proj.bias [3 072]
model.diffusion_model.double_blocks.0.img_attn.proj.weight [3 072, 3 072]
model.diffusion_model.double_blocks.0.img_attn.qkv.bias [9 216]
model.diffusion_model.double_blocks.0.img_attn.qkv.weight [9 216, 3 072]
@yilingsummer837 It could be the approach I used to merge this latest model is not concating the blocks correctly
@Felldude Thank you, I want to train a LoRA against this to make it generate 1 face consistently. I load this model with FluxTransformer2DModel.from_pretrained but it doesn't work because the state dict key from your safetensor is different. Do you have any suggestions?
@yilingsummer837 Save the model only using comfy, if the model still won't load with default TE, do a concat merge of the model with base FLUX
this will take down any of regular schnell capabilities ?
It can still make an image at 2 steps, 4 steps etc
@Felldude about the objects or composition that can do , your train affect that ? (reduce any of original capabilities of the original model?)
@amazingbeauty It passes most logic test, red cube on green ball on yellow table - it appears your translating using an app maybe post in your original language
@Felldude 'passes most logic test, red cube on green ball on yellow table' this is what i meant , that your train not affect the original capabilites of flux , thanks .
if you will upload gguf q8 of your v7 (or latest updates) would be good.
if you will upload gguf q8 of your v7 (or latest updates) would be good.
Hey, can you provide a separate link for the workflow for the nf4 model, I am unable to load the workflow from images. Thanks :)
The NF4 images are in Forge rather then Comfy - do to the lora support in forge