fp8 needed for 24gb
fp8 needed for 24gb
Fcompare checkpoints
Gi left it on
Gtesting to see the difference after i did the speed boost, i didnt do that last time and it took 14hrs
Gdid 32 images too
Gthis time doing 21 images
DWindows version is still running slow?
Fyep
Fi opened 2 threads on diffusers and pytorch
Fno replies yet
F
Foverfit but details - 1536 px - 256 lora rank - fine tuning will be a blast
F
F1024 vs 1536 - so much more details captured
B
BMay I ask why the descriptor breaks off in the middle and suddenly stops
Ihi guys 
Bhi
Iwill try this training 
II was about to ask about the captions, 300 its enough? or it needs more?
SMight be because of max_tokens set to 300
Ihow many tokens its acceptable?
Bthx,let me try again
B1000
Ikkz thx
B
BEven 1,000. It's not even written.
SNow that's what you call a caption
SCould write a whole book
FMy first attempt training a Flux LoRA
Iwhats the subject?
FLady from DMC
Ii c
FData set was 43 images or so based on 4k screenshots of the game I scaled down to 1024x1024
FI feel I'm in uncharted waters as I'm so used to training SDXL models and generating images with a1111
FI'm using comfyui right now and I think I hate it
Fincrease max new tokens
Fuse swarmui :d
Fi am gonna train a community lora with unsplash dataset i have
F10400 images
Fwill caption with joycaption and share dataset on patreon
Factually it will be 20800 images since i will include bucketing version as well
Ijesus
PHey @Dr. Furkan Gözükara 
I’m trying to install CUDA 11.8 on RTX 4090 for training flux.
Every time I try to install it fails. I’ve tried a couple of different versions but still no luck.
I’m using windows 11 pro, with Nvidia studio driver version 531.41
I’m trying to install CUDA 11.8 on RTX 4090 for training flux.
Every time I try to install it fails. I’ve tried a couple of different versions but still no luck.
I’m using windows 11 pro, with Nvidia studio driver version 531.41
AI myself have glasses and trained a lora with 20/25 images with glasses, it always puts them by default. Have not used captions tho because I trained using ai-toolkit on replicate.
One issue I have isit always puts my face on other people in the background too and is heavily influenced by the background I had on the input training images (old city buildings)
One issue I have isit always puts my face on other people in the background too and is heavily influenced by the background I had on the input training images (old city buildings)
K@Dr. Furkan Gözükara can you please add the option in joy caption to cut off the last sentence. It’s always cut off at the end. So and option to cut of the last one would be perfect.
Ninteresting. maybe a bit overtrained then? you could try a version of the lora based off fewer epochs\steps if ai-toolkit saves them along the way? or specify a background in your prompt?
AYes totally, when specifying backroung in prompt it does work as intended
AI used 2200 steps in training v2, and 1000 for v1
ABoth are the same in results tbh
SDoes anyone have different/worse results with Flux LoRA’s in Comfy than in Forge for example?
