Software Engineering Courses (SECourses)•3y ago

Any idea what this error is

Tthewingag ValueError: Cannot load <class 'magicanimate.models.appearance_encoder.Appearanc...

thewingag•12/8/23, 2:35 AM

@Dr. Furkan Gözükara hjx, can u help me

Pelolisu•12/8/23, 3:10 AM

Pelolisu•12/8/23, 3:12 AM

Anyone having trouble using the provided class files? my Dreambooth tries to generate the class even though it already exists in the directory. I think its occouring only class 'man' for a 'girl' its ok.

FFurkan Gözükara SECourses cant tell need research

DR.Siva•12/8/23, 8:27 AM

Need 512*512 Lora training setting for animatediff.
We trained our faces for sdxl and dreambooth..

hozaggers•12/8/23, 10:14 AM

is there any advice for training in kaggle when your input images are a variety of dimensions? A lot of my training images are larger than 512512 but smaller than 10241024, and not necesarily square. Resizing them can give them unrealistic/stretched proportions, not sure if that might impact the training? or is it not possible to train these?

Hhozaggers is there any advice for training in kaggle when your input images are a variety ...

Deleted User•12/8/23, 10:16 AM

There is image resize script on the patreon

hozaggers•12/8/23, 10:20 AM

but question is will resizing help if the proportions are all distorted?

hozaggers•12/8/23, 10:21 AM

like taking a 600x800 picture and resizing to 1024x1024 will require different stretching proportions and it can make a full body shot look strange when you force it into that size

Tashiketh•12/8/23, 10:39 AM

You can also just use birme.net. Way easier than command line

AAIGambino Any idea what this error is

Furkan Gözükara SECourses•12/8/23, 10:40 AM

install error. i would a fresh install. you can also manually install bitsandbytes

Tthewingag @Dr. Furkan Gözükara hjx, can u help me

Furkan Gözükara SECourses•12/8/23, 10:40 AM

hi what help you need?

Furkan Gözükara SECourses•12/8/23, 10:40 AM

ah with comfy

Furkan Gözükara SECourses•12/8/23, 10:40 AM

sadly i cant help

Furkan Gözükara SECourses•12/8/23, 10:40 AM

i have standalone installer that works

PPelolisu Anyone having trouble using the provided class files? my Dreambooth tries to gen...

Furkan Gözükara SECourses•12/8/23, 10:40 AM

yes happens when mismatch of resolution

Furkan Gözükara SECourses•12/8/23, 10:41 AM

either training images or reg images res mismatching

DDR.Siva Need 512*512 Lora training setting for animatediff. We trained our faces for sdx...

Furkan Gözükara SECourses•12/8/23, 10:41 AM

makes sense. i should research it

Hhozaggers is there any advice for training in kaggle when your input images are a variety ...

Furkan Gözükara SECourses•12/8/23, 10:41 AM

you can try both of these

Furkan Gözükara SECourses•12/8/23, 10:41 AM

enable bucketing

Furkan Gözükara SECourses•12/8/23, 10:41 AM

without enable bucketing

Hhozaggers but question is will resizing help if the proportions are all distorted?

Furkan Gözükara SECourses•12/8/23, 10:42 AM

your only chance is first upscale and then crop. or as i said use bucketing and give it a try

FFurkan Gözükara SECourses makes sense. i should research it

DR.Siva•12/8/23, 10:43 AM

Wish there was way to convert /downgrade lora.. But I think it will only take half time to train Lora. With 512 trained Lora, we can use with animatediff. Also by training it, we could have trained for all three models

FFurkan Gözükara SECourses i have standalone installer that works

thewingag•12/8/23, 10:44 AM

so sad. thanks

DDR.Siva Wish there was way to convert /downgrade lora.. But I think it will only take ha...

Furkan Gözükara SECourses•12/8/23, 10:45 AM

LoRA training is not that fast

Furkan Gözükara SECourses•12/8/23, 10:45 AM

it is like 1.3 second / it vs 1.6 second / it

Tthewingag so sad. thanks

Furkan Gözükara SECourses•12/8/23, 10:45 AM

ye comfy is hard

FFurkan Gözükara SECourses your only chance is first upscale and then crop. or as i said use bucketing and ...

hozaggers•12/8/23, 10:53 AM

I'm thinking upscale and crop may still pose a problem if the aspect ratio is not square on the original image. I'm thinking I may need to sacrifice some detail/body on those in order to get a square fit? Does your resizer script help with that or will it not cut out body parts to make it square?

Furkan Gözükara SECourses•12/8/23, 11:01 AM

i have 2 scripts

Furkan Gözükara SECourses•12/8/23, 11:01 AM

when you run both

Furkan Gözükara SECourses•12/8/23, 11:01 AM

you will get square

Furkan Gözükara SECourses•12/8/23, 11:01 AM

with best quality

hozaggers•12/8/23, 11:04 AM

thanks, will I need to upscale first to above 1024x1024 if the images are smaller than that size before running the scripts?

mikemenders•12/8/23, 12:38 PM

I would also like to say a few words about the sample photos. I am currently training a Lora, where all the training photos are upper body shots. My training burned very fast and I just went down to a d_coef of 0.65 to get a good lora. So that's why it's good to take your sample shots from as many angles and distances as possible, because you can make a stronger and more flexible Lora. Just taking the same type of photos will quickly get stuck and the model will learn too much.

Zono50•12/8/23, 1:26 PM

https://github.com/havenhq/mamba-chat

GitHub

GitHub - havenhq/mamba-chat: Mamba-Chat: A chat LLM based on the st...

Mamba-Chat: A chat LLM based on the state-space model architecture

GitHub - havenhq/mamba-chat: Mamba-Chat: A chat LLM based on the state-space model architecture

FFurkan Gözükara SECourses install error. i would a fresh install. you can also manually install bitsandbyt...

AIGambinoOP•12/8/23, 2:09 PM

it works fine with sdxl but not regular sd1.5 training, weird

midas•12/8/23, 2:22 PM

Hey guys

FFurkan Gözükara SECourses either training images or reg images res mismatching

Pelolisu•12/8/23, 2:52 PM

my provided dataset is 768x1024 for realistic and image processing is 1024 in max resolution parameter. I followed the tutorial, first I adjusted the aspect ratio with the cropped.py script then I pre-processed the image at 768x1024 in the train tab. I have provided class files in the same resolution, however, the images it generates in the directory are 880x1176. I think I'll have to use it at 512x512 with realistic_vision_v5.1.

AAIGambino it works fine with sdxl but not regular sd1.5 training, weird

Pelolisu•12/8/23, 2:54 PM

I gave up on Windows, I'm using the latest version of Dreambooth on WSL 2 Ubuntu.

FFurkan Gözükara SECourses yes happens when mismatch of resolution

Pelolisu•12/8/23, 3:15 PM

You were right, the problem was that the dataset was 768x1024 and the classes were 1024x768, I adjusted the dataset to 1024x768 and now it has gone to the cache latents phase. I think it will work now.

Tthewingag ValueError: Cannot load <class 'magicanimate.models.appearance_encoder.Appearanc...

Pelolisu•12/8/23, 3:26 PM

I haven't used ComfyUI yet, but the error is similar to the dreambooth save preview. If it uses transforms it could be the problem with the new versions in Torch 2.0. Try using an old version that used Torch before 2.0. Then if I have time and find out something about it I can share it. I have fixed many issues in my local version of dreambooth.

PPelolisu I haven't used ComfyUI yet, but the error is similar to the dreambooth save prev...

thewingag•12/8/23, 3:47 PM

thanks. i fixed error but i has got another error. lol

thewingag•12/8/23, 3:53 PM

ValueError: Cannot load F:\AI\ComfyUI_windows_portable\ComfyUI\models\MagicAnimate\control_v11p_sd15_openpose because down_blocks.0.attentions.0.transformer_blocks.0.attn2.to_k.weight expected shape tensor(..., device='meta', size=(320, 1280)), but got torch.Size([320, 768]). If you want to instead overwrite randomly initialized weights, please make sure to pass both

low_cpu_mem_usage=False

low_cpu_mem_usage=False

and

ignore_mismatched_sizes=True

ignore_mismatched_sizes=True

. For more information, see also: https://github.com/huggingface/diffusers/issues/1619#issuecomment-1345604389 as an example.

GitHub

Adding additional input channels to model after intialization / Con...

Have scoured the docs for an answer to this, to no avail. Is it possible to add additional input channels to a model after initializing it using .from_pretrained. For example (taken from your Dream...