yes of course, I mean in your video its was Dreambooth and Dreambooth lora ( exact name) maybe renam
yes of course, I mean in your video its was Dreambooth and Dreambooth lora ( exact name) maybe renamed to lora only in the newest version .
Fthey renamed
MI have now tried the Prodigy Lora training with another parameter that is also recommended for SDXL. In general, in both cases, epoch 10 (the last one) gives (for me) absolutely satisfactory and stable results.
- This is the optimizer args that I am using now: "use_bias_correction=True" "weight_decay=0.01" "decouple=True" "betas=(0.9,0.999)" "d_coef=2" "safeguard_warmup=False"
It makes the model more flexible and works well on Clip Skip 1. The way I look at it, 3-6 epochs already give good results, at higher epochs you can see one of the characteristics of overdrawing, the freckling.
- this is what I used yesterday: "use_bias_correction=True" "weight_decay=0.5" "decouple=True" "betas=(0.9, 0.99)" "d_coef=2" "safeguard_warmup=False" (here the weight_decay value is stronger)
This is also very good from 3, and even 10 didn't burn. But this gives an absolutely strong, realistic model even at 512x512 resolution, without Hires Fix. This is worth training on Clip Skip 2 to make it flexible. I got such a perfect result that it not only restored the tattoo to its position, but even its shape. However, it does not like getting naked because of this.
So if you want very realistic images, the latter is better, if you want more flexible, the former.
Update1: Epoch 4 seems to be the most flexible out of the 10, after that you need stronger prompts if you want to see not only the face but also the body.
- This is the optimizer args that I am using now: "use_bias_correction=True" "weight_decay=0.01" "decouple=True" "betas=(0.9,0.999)" "d_coef=2" "safeguard_warmup=False"
It makes the model more flexible and works well on Clip Skip 1. The way I look at it, 3-6 epochs already give good results, at higher epochs you can see one of the characteristics of overdrawing, the freckling.
- this is what I used yesterday: "use_bias_correction=True" "weight_decay=0.5" "decouple=True" "betas=(0.9, 0.99)" "d_coef=2" "safeguard_warmup=False" (here the weight_decay value is stronger)
This is also very good from 3, and even 10 didn't burn. But this gives an absolutely strong, realistic model even at 512x512 resolution, without Hires Fix. This is worth training on Clip Skip 2 to make it flexible. I got such a perfect result that it not only restored the tattoo to its position, but even its shape. However, it does not like getting naked because of this.
So if you want very realistic images, the latter is better, if you want more flexible, the former.
Update1: Epoch 4 seems to be the most flexible out of the 10, after that you need stronger prompts if you want to see not only the face but also the body.
Fjust found out amazing thing
Fwarp of cloudflare
Ffixes pip install network issues haha
Ffree vpn
EAnyone have any idea why runpodctl slows down when it arrived at 95% completion? i get 80mb/sec at start, 14mb/sec after 20% and its gradually slowing down as time progresses down to 100kb/sec
Fye it is really slow
Fnow i prefer hugging face
EI got this bad resaults in dreambooth training, did 8000 steps, 24 images, 42 repeats, 8 epochs
Why did the face come out liek this ? (happens in all generations)
Why did the face come out liek this ? (happens in all generations)
Seven on closer up renders?
Eim using Adetailer for face, but ill try close up
Eyes even on closeups
Swhat is your cfg at?
E7
Stry 3 or 4
Si never used auto detailer so not sure on that but sometimes cfg needs to be lower
Fdid you try previous epochs?
Fi am working on an auto captioner best one
Eyes, i saved every 2nd epoch so i have 2, 4, 6, 8
but all of them makes this kinds of weird bugs
but all of them makes this kinds of weird bugs
FViT-bigG-14', 'laion2b_s39b_b160k
Fdm me your print training command
EI get this two photos when i try "photo of ohwx man" cfg 5 no adtailer
Fsomething wrong with your training possibly
Esent you the command on dm
F1 more thing
Fcan you test with
FDPM++ 2M SDE Karras
F40 steps
AI have 8 gb VRAM:)
And this settings really helped for me
And this settings really helped for me
Fnice
Fyes --medvram super important
Feven 12 gb need it for sdxl
AYeah, I compared the SDXL and SD 1.5 images and really the improvements are not worth the increased system requirements for comfortable use. I think it will be at least another year before SDXL is would be fully usable by most people.:)
Fwell i relly find SDXL superior
Mone area I realized with training on 48 gb vram and using models on 3060 12gb vram was that high quality may have been trained into the model but you can't access (recreate) it with lower vram systems. My trained models have come to life with an upgrade to 4090 24gb vram. I didn't think the difference would be so noticeable. SDXL has much improved prompt response and much higher quality. I recommend that if someone wants both high quality, flexibility and speed runpod or kaggle is the best bet. I wouldn't recommend going back to 1.5 - granted it is an expensive upgrade..
MAnother comparison: Adafactor vs Prodigy.
The Adafactor is slower to train and holds the prompt better on these lora. The lora trained with Prodigy is strong and the training is fast, (4th epoch seems to be good if it's just a face). After that it picks up other features (hair, body shape, tattoos, etc.)
So for a person I prefer Adafactor, for a style or concept I prefer Prodigy, but if you need a realistic person you can use Prodigy for that too, but you may need Clip Skip 2 or 3 for flexibility when generating person images, because Clip Skip 1 is just an image generator, nothing more.
The Adafactor is slower to train and holds the prompt better on these lora. The lora trained with Prodigy is strong and the training is fast, (4th epoch seems to be good if it's just a face). After that it picks up other features (hair, body shape, tattoos, etc.)
So for a person I prefer Adafactor, for a style or concept I prefer Prodigy, but if you need a realistic person you can use Prodigy for that too, but you may need Clip Skip 2 or 3 for flexibility when generating person images, because Clip Skip 1 is just an image generator, nothing more.
FI am preparing SOTA image captioners 
Fauto install auto run so easy
PSOTA?
Famazing script coming
FSOTA means state of the art
Pahh. thought it was a model I'd never heard of, lol. I've been looking at blip models to do captioning. but not sure which one would be optimal. it's really the main reason I haven't gotten into LoRA training on sdxl. don't want to spend hours captioning when I can get ai to do it. then all I'd have to do is go back and make sure the captions align with what I'm wanting
Fi am preparing best ones
Fblip2
Psweet. yeah, I was looking at the blip2 models on the salesforce huggingface. but sooo many models