SDXL model please

[ynie]

Hello, I would love to use this in my SDXL workflow. Thank you for your work.

[jshilong]

Since it was originally just a personal research project, we didn't utilize a significant amount of computational resources. However, if there is a higher demand for SDXL, we might consider using a larger model to achieve better results.

[akk-123]

@jshilong can flashface use other community sd-1.5 as base model?

[jshilong]

@jshilong can flashface use other community sd-1.5 as base model?

No, because it is fully finetune

[3blackbar]

Hey man This is the best method for recreating likeness from photos, please please make SDXL version !
I have instantID, tried ipadapters, this one is the best one .
Lot of people did not tried it yet cause there was no comfyui nodes but now they exist and i run this for hours to genereate pics of me as a kid and it looks pretty amazing

[jshilong]

Hey man This is the best method for recreating likeness from photos, please please make SDXL version ! I have instantID, tried ipadapters, this one is the best one . Lot of people did not tried it yet cause there was no comfyui nodes but now they exist and i run this for hours to genereate pics of me as a kid and it looks pretty amazing

The usage of FlashFace requires frequent parameter adjustments to balance language control and character fidelity, which has been a source of frustration for me. It is not user-friendly, and I have always had doubts about whether users would appreciate it. Therefore, receiving your positive feedback has made me very happy.

[3blackbar]

From what i see its literally pasting the reference face into a picture, its why it has the best results of them all, with SDXL it would gain great quality especially with new models like dreamshaperXL_v2TurboDpmppSDE which is the best one or nightvisionXL, they bot h get crips sharp results and dont require 2nd model for details anymore unlike stock SDXL 1.0 which i DO NOT recommend to train on.
Im able to get 1024 output from flashface but not sure if im already stretching sd1.5 too much with it
Are You able to unpack the weights of flashface and swap the VAE ? there is new one that has better eye details - vae-ft-mse-840000-ema-pruned.vae.pt https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.ckpt

Is the flashface in UNET or text encoder and vae? can we swap any of these to other models ? Did anyone tried ? Was flashface trained on stock sd1.5 ? I think this model is not pruned to fp16 ? This is trained on 4GB sd1.5 model and was not pruned to 2GB.
Absolute best results with flashface im getting with bbox1x y1 on 0.1 and bbox x2 y2 on 0.9. Cause the face fills whole frame ,likeness its identical to reference. with just 6 pics
I dont have many pics of me from childhood and this tool made it possible , i now have a lot and its pretty amazing to see , same with people close to us that died and we dont have more photos of them, we can bring them back for a moment.