If you like what I do, leave me a ❤️ to receive notifications of future updates and follow me here on Civitai. Remember that it is a hobby for me and it costs a lot of time and everything is shared entirely for free, I appreciate that you make a review and share your creations on this wall.
Project Gaia Flux1.D V2 NF4 FP8-T5XXXFP8 Uncensored
Project Gaia Flux1.D V2 NF4 FP8-T5XXXFP8
New Data Injected
Uncensored
less vram consumption
Project Gaia Flux1 V1
Here we are altering the many layers of this fantastic model that Black Forest Labs has gifted us.
In this first attempt, the Flux 1 Dev fp8 model has been used as a base. It has then been merged with my LoRa: Image Enhancer Flux By Devildonia (soft version) and then converted into NF4 by incorporating the vae along with the t5xxxl fp16 clip and the L clip.
This V1 has no pretensions other than to learn from this new model, alter the original enough to obtain something new of quality, and lay the foundations for future updates.
I hope you like it, and see you in the next update.
P.D.
Clips and VAE included. If anyone is interested in just the UNET version, without the clips, let me know.
For the Unet version, you need to download VAE, CLIP L and CLIP T5XXL separately for it to work.
Description
v2.0 NF4 UNET ONLY Uncensored
FAQ
Comments (12)
Oh wow, works perfect and very fast on WebUI Forge,- on different sampling methods and schedule types, CFG Scale sweet spot is ~5ish. Definitely thumbs up😎
loras react yes, but unfortunately portraits of people, for example, all have to be retrained, they looks overtrained and if you reduce its not the face anymore
You are right, I made the mistake of training V2 from V1 which was mixed from several LoRa, that results in overtraining. I am currently retraining everything from scratch with the V2 dataset and already in the preliminary tests it gives much better results.
Can you add the GGUF version too?
Im with that, cause Lora is no workie for me with NF4.
pls release FP16 or FP8 not nf4... as macs don't support nf4 (no bitsandbytes support)
现在NF4已经是落后的（NF4节点作者似乎已经放弃更新并建议使用GGUF），建议增加GGUF版本
Just a question, output images have in my case barely but visible grid/mesh lines not matter what I try, if I just switch to original Flux (same parameters, same seed) then same picture is clean... what is the cause ?
nf4 is a type of compression and it affects the final result. Playing with the cfg it is possible to make those lines not appear. I imagine you use forge, in comfyui those lines do not usually appear.
@DeViLDoNia ok, I usually use Forge but I also have Comfy, I can have a try if it happen there too, I have both NF4 and normal version of Flux and your Project and tested both and seems with original flux doesn't happen but ok, I go test more :)
I'm using a LoRA meant for Flux 1.D. It works well on flux1DFp8 but when I use it on your model it doesn't look right. I'm new to flux, is this model not compatible with the LoRA?
Never mind I discovered I cant use it in COMFYUI. Do you have a comfyUI version?