Is it even possible to set batch size 7 with VRAM 48GB? Because even with batch size 1, VRAM usage i
Is it even possible to set batch size 7 with VRAM 48GB? Because even with batch size 1, VRAM usage is around 36GB
MFYI, I want best quality, but I should also consider training speed
Fi think not with lora
Fit is for dreambooth at 1024x1024
MOh, I see
MFor dreambooth, batch size 1 & 7 gpu vs batch size 7 & 1 gpu
MIs this same speed?
MI assume quality would be better with batch size 1
Fof course not
Fnormally 7 gpu means it scales
Fwith lora almost 7x speed
Fbut for dreambooth you need 80 gb gpus and i didnt test
Fbatch size 7 is like 1.3x speed
MThen, best way to speed up the training is
- lora
- batch size 1
- multiple gpus(for linear speed up)
MAdditional question, is there any way that I can improve training speed itself?
- I used best Tier 1 config for 48GB vram for lora training
- But, set "cache_latents = false / cache_latents_to_disk = false"
- With RTX 6000 Ada, it took 24 mins for 20 epoch
- With A100 SXM, it took 16 mins for 20 epoch
Fsadly not much else
Fgraident checkpointing slows down
Fbut turning it of requires huge vram
Mhmmm... in that case, I need to think about how to optimize gpu cost
MMI'm curious how much quality difference between fp8 and fp16
Fnot possible
Feven if you use fp8 weights the training happens at 16 bit precision
Fkohya doesnt support fp8 training
MThank you so much for answering all the questions! I'll focus on facial accuracy with Kohya fisrt, and then think about speed later haha
ZI see, thank you.
Where we can see the class type for this flux model?
Where we can see the class type for this flux model?
Fthere is no class types
Fyou need to understand yourself sadly
HYoo devs
U@Furkan Gözükara SECourses Hello 
I was wondering if you have any plans in the future to train a custom lora for wan 2.1/2.2 (when it releases) video generation?
I was wondering if you have any plans in the future to train a custom lora for wan 2.1/2.2 (when it releases) video generation?
Fi saw today wan 2.2 coming haha
UCant wait for your video 
UWhen should I be looking out for it?
Fafter this video upscaler i plan to research them :d
Fi need to fisnih this
UCan't wait
UHey, I have been finding resources and documents for text to video for wan with images being dataset or image to video with videos being dataset, but I was wondering if there is a way to train lora with images and use first and last image as reference and generate a video based on the reference images? do you happen to know any models which lets you train on images and then use that lora with reference images?
Fi dont know such model
Fbut wan 2.1 can be trained from images to generate videos
0Hey guys, so I'm really impressed with the fine tune accuracy with regards to my model's face. it looks great. but it still has a bit of AI look to it. and I know the model is trained on photorealism.
my goal is it make it look more candid, like it was shot on an iPhone or a digital cam. I've used loras with varying strengths and it completely changes the faces on the model. I've tried different resolutions but it doesnt help much.
my goal is it make it look more candid, like it was shot on an iPhone or a digital cam. I've used loras with varying strengths and it completely changes the faces on the model. I've tried different resolutions but it doesnt help much.
0should I try and combine the fine tune with a realism fine tune or would that totally mess it up?
Fmore realism is obtainerd with
Fmore images + more training
Fbut if you use a more realism oriented fine tuned base flux model
Fit may also help
0I actually just tried that but it keeps giving me an error. Specifically I'm using ultrarealfinetunev4 as the base model and its saying
KeyError: 'time_embed.0.weight'
Chatgpt is saying:
"You're running into a common issue: the UltraRealFineTune .safetensors file is not a full Stable Diffusion checkpoint — it's missing key components, particularly UNet time embedding weights, which causes this error:
KeyError: 'time_embed.0.weight'"
KeyError: 'time_embed.0.weight'
Chatgpt is saying:
"You're running into a common issue: the UltraRealFineTune .safetensors file is not a full Stable Diffusion checkpoint — it's missing key components, particularly UNet time embedding weights, which causes this error:
KeyError: 'time_embed.0.weight'"
0so assuming I cannot use that model do you have any reccomendations?
Fchat gpt or alikes will know nothing
Fare you training stable diffusion
