are you using latest version?
are you using latest version?
Fof kohya
DI cant remenber what was diferent
Dit was rank 8
Dnow is working rank 8
D
Dfor Network Rank (Dimension) 8 I need to set Convolution Rank (Dimension) to 8 too?
Dand what about Network Alpha and Convolution Alpha?
Dyes I did a git pull before training
DI'm going to test your lowest_vram config now
DNow is working with your 1e_04_fp8_bf16_accelerate_full_bf16_32_rank config, I just restarted the PC and enabled buckets not all my images are 1024x1024 is working with great speed
Dthat was with my monitors turned off but I need to use my pc for other tasks, using the pc the speed decreased
�Is 200 epochs per image really necessary tho? 387 images in your dataset and training will be finished like in 4 days lol 
Dhehe yes
DI can stop it earlier and test
Dbut yes is too slow
![Dazzastrous [4090]](https://cdn.discordapp.com/avatars/871389071686135838/c4a5f4d81d9eeb18d856295d692e2c07.webp?size=40)
Dany idea why my comp completely freezes when trying to install Dlib in conda?
FConvolution Rank is for other lora types as far as i know
Fmake network rank 8 and alpha 8
Fnope i am doing big test
Fi will aim 100-150 epoch
Fi have 15 classic dataset
![Dazzastrous [4090]](https://cdn.discordapp.com/avatars/871389071686135838/c4a5f4d81d9eeb18d856295d692e2c07.webp?size=16)
Ftesting newest loras
F
F
Fno face inpainting testing different results - no cherry pick
Vwhat is the current captioning/image interrogator darling thats most popular this month?
Fflorance is said to be good but i havent had chance to make a graido yet
FGitHub
Contribute to wegwerfen/flocap development by creating an account on GitHub.
VDoc, is crack back in Turkey? You've been dropping alot this month. lol
Fi was always in Turkey 
Fi am still at mountain house though
Fno face inpainting i hope we never be ahve to anymore :d
Vnice. best of both worlds, matrix ai and fresh air and sky
Fyep only issue is internet
Fit is quota based and terrible
Vthats a bummer.
Vmaybe renting a remote win10 VM in some Amazon datacenter?..
Vanyway, enough small talk. ive got alot of homework going through your August tuts. I'll check out Florence.
Fwell i am usng massed compute but dont have rmeote for windows
Fhaha 200 epoch high LR completely memorized
FI love the queue system of SwarmUI very much 
F6e05 is really good but requires 200 epoch training :/
�Florence is good but isn't as accurate as I'd hope with certain things... had to make fixes to a lot of captions. I'm still looking for something that feels more automated. This looks promising though: https://huggingface.co/spaces/fancyfeast/joy-caption-pre-alpha
VNice. thanks for sharing. I'll check it out and test it. Overall I'm seeking local/self-hosted and open/ungated/uncensored products. Found some safety features tend to mute outputs. Qwen 2 I found to be very nice but inconsistent. Regardless it delivered wow factors and that's my baseline with certain models. In the ever changing landscape I feel like i gotta ask every month to see who's on top since nerfing is a thing for the cloud stuff atleast.
�Yeah same here, I'm looking for something Local too. You can actually git clone this repo and then modify the script to be local instead: https://huggingface.co/spaces/fancyfeast/joy-caption-pre-alpha/tree/main
�I saw these instructions, but I haven't gotten around to messing with this yet
1. git clone the linked repo, it has the trained clip-to-llama thingamajig included already
2. setup the venv, or if you are lazy like me, use an existing venv from one of the other billion projects with a superset of this one's requirements.txt
3. edit app.py to remove all the gradio and spaces junk
4. replace "meta-llama/Meta-Llama-3.1-8B" with "unsloth/Meta-Llama-3.1-8B-bnb-4bit" to save space and not have to authenticate with hf
5. print(stream_chat(Image.open("/path/to/boobas.png")))
6. ???
7. profit