HII Create your own sliders at https://sliders.ntcai.xyz
-1 is sad and 1 is happy
Only A1111 works correctly. ComfyUI has an open issue with DoRA support here https://github.com/comfyanonymous/ComfyUI/pull/3298 Add a reaction or comment if you need this to work in comfyui.
Trained for 6000 instances with sliders https://github.com/rohitgandikota/sliders modified for DoRA
batch size 12, 500 steps, 1.5 hr on 1x a6000
Trained on https://civarchive.com/models/384401?modelVersionId=429011 . New evolved SDXL lightning model.
Support my work on patreon and get access to over 1000 unseen loras and 10,000 lora merges (more daily) at https://sliders.ntcai.xyz! Plus, easily train your own - no hardware needed!
Any proceeds go to more gpus - support what you enjoy and enable my addiction! See screenshots below.
Patreons get (NEW): Easily create your own private sliders with text alone.
Patreons get: Combined Loras result in unique styles. Get access to thousands and create your own
"Theres something coming"
Description
FAQ
Comments (22)
I Love all of those sliders. If you ever release freely the eye (open/close) slider. I'd gladly donate buzz.
Thanks! I get to discover what the terms do more than direct them. You can experiment with different terms at sliders.ntcai.xyz but it costs money for the hardware so it's $5 / month. No worries if you can't or don't want to.
too much biases, what dataset you used bro, almost only nsfw naked people
The model I trained on and the technique are stated in the description - it doesn't use a dataset.
Not sure about the nsfw things? Maybe those people are happy?
@ntc no matter if it's happy or sad, it produces naked people with any weight, no matter what i put in negative prompt
@rns96 This is not my experience. I use DoRA is dev branch of a1111 with the toprated evolved model above.
At zero strength there should be no changes applied.
@ntc maybe something related to attach DoRA in other models, making them to show more biases. Maybe something about DoRA structure itself, but could be the training as well, is something very new, very experimental.
@rns96 What SDXL software did you use and what branch? And what model?
@ntc ComfyUI latest, also tried in the checkpoints ZaxiousXL lightning and nightvisionXL
How is this a problem again? lol
What did you think about the effectiveness of DoRA vs LoRA?
Subjectively DoRA seems like it has more modes. I suspect it's a straight upgrade and I plan to switch the sliders to it once the tooling catches up.
@ntc Good to know. I guess I should get the team moving towards that transition as well then.
@JustMaier Nice, if they are curious this is how I implemented it https://github.com/ntc-ai/sliders-conceptmod/blob/main/conceptmod/textsliders/dora.py#L143
Wow I was just wondering if this would ever be implemented into Civitai and I guess my question has been answered😂 excited to see how this will turn out, hopefully it reduces your gpu costs if anything
What is the most important reason for us to switch from Lora to Dora?
I believe it listens better to what you want, and if using a dreambooth or something changes the underlying subject a lot less and just general improvement.
@user1234123 @samliu Yeah it should just work far better for the same compute cost, especially at lower ranks (lower file size). https://sebastianraschka.com/blog/2024/lora-dora.html is an amazing blog post about it.
Any chance you can share how you trained a Dora? I tried Onetrainer but it doesn't seem like it's supported yet.
Yeah I wrote a dora module then trained it with sliders and tested with a1111 dev branch https://github.com/ntc-ai/sliders-conceptmod/blob/main/conceptmod/textsliders/dora.py
Hey
In the github implementation of dora.py you have shown, how exactly do we train the model, where and with what should we replace the dora.py code? Trying to run it and I am getting a little confused. Please assist
Hey, heres the command I use to train
CUDA_VISIBLE_DEVICES=0 python train_lora_xl.py --name 'happy' --rank 8 --alpha 1 --config_file <configfile> --peft_type dora