experiment model
This version only add embbedings from: XD
https://civarchive.com/models/2673845/realism-lighting
https://civarchive.com/models/2673866/realism-render
https://civarchive.com/models/2673798/realism-qcc
in here:
i didn't train the base model or merge to any model, I just tuned the base model by shifting it backward and increasing the some weight value 🤫.
CallEmXL-amateur-alpha2
1. add hdr + 8k filter..
2. continue remove all glamour concept..
3. now i try inject to unet :D .. )
4. the model forget to iron the clothes :D
5. the right prompt now how to ask the model to doing something and not doing pose like something :)
==================
bug:. that because using HDR + 8k it make eye pop out and face like brick on medium shot ++.. still searching the right channel for that T_T
CallEmXL-amateur-alpha1
1. mean remove all glamour professional concept and pose,
2. model will not always looking to straight to camera yah ,
3. add 255 basic color pallet :D if necesary
4. 92% vocab weight controlable by Juggernaut, using EOS on end .. the last model using standard
5. etc..
work with any lora lighting step like..
https://civarchive.com/models/350450/sdxl-lightning-loras
https://civarchive.com/models/1608870/dmd2-speed-lora-sdxl-pony-illustrious
vae:
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl.vae.safetensors
Personal/Non-Commercial only. Commercial use requires written permission.
This model is a modification based on
Juggernaut XL (Ragnarok) created by RunDiffusion.
All credit for the base model goes to them. 🔗
Original model: [https://civarchive.com/models/133005/juggernaut-xl]
Description
FAQ
Comments (5)
This is really interesting to see TI / embeddings merged into a checkpoint. I've been thinking of experimenting with this for a while. Did you use a script to save the weights directly to the CLIP in comfyui?
yip on clipG&L, As usual... I've seen the flow from the initial prompt to the end. Usually, ti will remain in the token_embedding.weight layer, I thought, why not just inject ti and distribute the weights to all vocabulary vectors? I spent a month working on this while learning... so no training is needed... it all depends on the base model, to many injecting new token will make your model like stone.. so need search token and weight on CLIP-L&G, then you can replace/add weight to the rarely used/targeted tokens with the weight tokens you have injected,, make this script alone XD..
why i do this? because here it is limited to 3. u must know about that XD,, ciyayooo thanks jie..
sure it's only a small effect? its bad man
just experiment dude 🤣. but ti success inject :p . try with tag c4ll3m
only 0.03 weight from primary tag ,, still experimenting .. gluck