This version is a merge with my lora real-lora.
https://civarchive.com/models/650205/real-lora
It's a Unet version without the clips and without the vae, you have to load them separately
With forge i use this setting:
Sampler: Euler
Scheduler: beta
Steps: 20
Distilled CFG Scale: 2
Description
Fp8 Unet version without clips and without vae
FAQ
Comments (17)
tyvm for uploading this without everything baked in <3
Yes! Double agree. Running on a 7900XTX and the emulation layers aren't great with VRAM. This model runs like a champ!
@jice Sorry for the silly question. I need the clip and VAE in order to run this checkpoint but I do not kwow how to get them. Could you please enlighten me??
Thanks in advance!
@atrts yes, this is only the "Unet". Designed to work with the same default text encoders and VAE that the normal/default Flux models use. Unet only version allows you to select the 8-bit or 16-bit CLIP version of the text encoder. Also much smaller file size!
Thanks a lot once again guys, I really appreciate it!
I am really close in generating a proper image, at the moment I'm getting a pitch black though :)
So I put the linked Hugging_face files under the (Forge GUI) subdirectories:
/models/VAE (diffusion_pytorch_model.safetensors)
/models/text_encoder (clip_l.safetensors, t5ccl_fp8_e4m3fn.safetensors)
and then via the GUI, I chose the model (Checkpoint) and ALL of the above .safetensors files (VAE/Text Encoder).
I also set the Sampling method: Euler, Schedule type: Beta, steps:20,
Distilled CFG Scale: 2
and CFG Scale: 1
But no luck other than a black generated .png!
NOTE: During generation of the image, I can see the image I prompted, but eventually it turns black and gets saved this way. :S
@atrts In the vae folder, put the file ae.safetensors and select it on forge
@jice worked! Thanks a lot! Now I'm generating really cool images :)
@jice One last question, please.
In my understanding, Creaprompt is a Comfy_UI extension for generating more creative random prompts during inference time.
How does this LoRA adapter (or more rigorously this merge of current base_model+LoRA) include Creaprompt IN this safetensor??
I apologise if this is a stupid question but I'm still trying to figure out things :) .
@atrts CreaPrompt is an extension for A1111/Forge/Comfyui and a Checkpoint (basically sdxl), the two are different. There is also my latest version Hyper-Flux-Dev 8 steps
https://civitai.com/models/699688?modelVersionId=789966
https://github.com/tritant/ComfyUI_CreaPrompt
https://github.com/tritant/sd-webui-creaprompt
https://civitai.com/models/383364/creapromptlightninghyper-sdxl
Very nice. I still use this with most of my images even after so many others have been released. I also appreciate that it is without the VAE, etc.
Thx for your feedback, you cant try my last version: https://civitai.com/models/699688?modelVersionId=930403
congratulations and thanks for this great model, the skin is pretty good and realistic. With forge it works well and also pairs well with fast flux lora 8 step. Are you thinking of making a FP16 version or a Hyper version? That would be great. Thanks.
Thx for your feedback, have fun
@jice oh yes, I know 8 steps CreArt-Hyper which is very good and I will continue to use it, but personally as a neophyte I think that this CreaPrompt-Flux.1-Dev is really fantastic and very good in the representation of the skin so much that I can do without additional loras to improve the skin. I don't know if it's only in my case that I use forge but the skin is really very realistic and I am very grateful to you for making it available. I know it's a hard work for you but I think that if you release others like this many people will appreciate it. I honestly think it's the best I've tried so far. Thank you very much.