you may try, but without them, my default shivam shiraro was unable to convert the model (the training was done till the end but in the dumping phase to diffusers it complained that those were missing)

I was actually surprised the first time because during the conversion of RV2.0 to diffusers those were missing, I had to take them from the previous model but fortunately, it worked for me

if you're going to remove them, you should also mind the json file (model_index.json) (either remove those two entries from it or set the values to null, but like i said - my repo needed it)

Haha, thank you. I hoped that someone would appreciate it :)

for A1111 you need ckpt/safetensors version which you can grab directly from here: https://civitai.com/models/4201/realistic-vision-v20

the diffusers version is used for other repositories like ShimamShrirao or well, diffusers :-)

instead of checkpoint format, those expect the diffusers format

@noneoofurbusiness can you clarify your question on what exactly is the main problem that you are facing?

I have attached the script parameters that I use to extract it from dreambooth into LyCORIS

pretty much what I do is execute the script from the command line as follows:

python D:/Development/StableDiffusion/LyCORIS/tools/extract_locon.py --safetensors --device=cuda --mode="fixed" --linear_dim=64 --conv_dim=32 C:/Development/StableDiffusion/stable-diffusion-webui/models/Stable-diffusion/realisticVisionV20_v20.ckpt F:/Development/StableDiffusion/SDModels/Models/converted/model_elizabethbanks_v1_realistic2_3000.ckpt C:/Development/StableDiffusion/stable-diffusion-webui/models/Stable-diffusion/locon/locon_elizabethbanks_v1_from_v1_64_32.safetensors

this:C:/Development/StableDiffusion/stable-diffusion-webui/models/Stable-diffusion/realisticVisionV20_v20.ckpt

is the path to the base model that you use for dreambooth training (in my case RV2.0)

this: F:/Development/StableDiffusion/SDModels/Models/converted/model_elizabethbanks_v1_realistic2_3000.ckpt
is the path to the created dreambooth model

and this: C:/Development/StableDiffusion/stable-diffusion-webui/models/Stable-diffusion/locon/locon_elizabethbanks_v1_from_v1_64_32.safetensors

is the output path where the LyCORIS model should be saved

you also see the dimentions params (64/32)

this is pretty much all in regards to extracting LyCORIS from the dreambooth :)

"i can't understand how he even changes model to realisticvision, i don't even know how to choose right path"

@UZY on the rentry page of the guide is the train_morgan.sh file

the first line is: export MODEL_NAME="/home/malcolmrey/models/realistic2"

you put there the path to your base model (in my case realistic vision2), the base folder needs to be in the diffusers format, if you want to use realistic vision 2 then in this guide there is the second version called base model in diffusers: https://civitai.com/models/45539?modelVersionId=81468

you need to get the zip and extract it to the folder that you put under the MODEL_NAME

@Stop_Civvy_Time unfortunately i think i know what could be the problem - it is most likely a mismatch of the libraries (transformers+diffusers or just one of those) between dreambooth and kohya

which extraction tool are you using? the standalone from kohakublueleaf or the one already included in the kohya_ss?

Hi @Fuckingdope

1. thanks! :-)

2. yes, shivamshirao does not have this switch: shuffle_after_epoch, it is exclusive to inb4 (so if you would like to try shivam in the future, just remove that parameter)

3. 1 hour and 30 minutes is usually my training times, i see that for you the training actually finished with 1 hour and 47 minutes: Steps: 100% 3000/3000 [1:46:57<00:00, 2.14s/it, loss=0.259, lr=1e-6]

seems that the error comes during the saving process itself, mainly this:

ValueError: The component <class 'transformers.models.clip.feature_extraction_clip.CLIPFeatureExtractor'> of <class 'diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline'> cannot be loaded as it does not seem to have any of the loading methods defined in {'ModelMixin': ['save_pretrained', 'from_pretrained'], 'SchedulerMixin': ['save_config', 'from_config'], 'DiffusionPipeline': ['save_pretrained', 'from_pretrained'], 'OnnxRuntimeModel': ['save_pretrained', 'from_pretrained'], 'PreTrainedTokenizer': ['save_pretrained', 'from_pretrained'], 'PreTrainedTokenizerFast': ['save_pretrained', 'from_pretrained'], 'PreTrainedModel': ['save_pretrained', 'from_pretrained'], 'FeatureExtractionMixin': ['save_pretrained', 'from_pretrained']}.

so my question is, what do you use as your base model? do you have a local model in diffusers format? common issue (at least for me) was that the diffuser format model was missing some folder (something was not provided in the safetensor/ckpt file that was not needed for inference but is needed for training)

just in case you do in fact use a local model, then in the folder of that model should be a model_index.json file with content like this:

{

"_class_name": "StableDiffusionPipeline",

"_diffusers_version": "0.10.2",

"feature_extractor": [

"transformers",

"CLIPFeatureExtractor"

],

"safety_checker": [

"stable_diffusion",

"StableDiffusionSafetyChecker"

],

"scheduler": [

"diffusers",

"DDIMScheduler"

],

"text_encoder": [

"transformers",

"CLIPTextModel"

],

"tokenizer": [

"transformers",

"CLIPTokenizer"

],

"unet": [

"diffusers",

"UNet2DConditionModel"

],

"vae": [

"diffusers",

"AutoencoderKL"

]

}

and you should also have a subfolder feature_extractor with preprocessor_config.json

I'm mentioning this specific folder because the error pointed to it: The component <class 'transformers.models.clip.feature_extraction_clip.CLIPFeatureExtractor'> of <class 'diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline'> cannot be loaded

I have provided in the second version 'base model in diffusers" so if you're missing some of the files, you can copy it from there, those are for the most part generic so they serve well as a replacement

also, if you want to test that this works fine, set steps to something low like 5 or 10 so that you can go to the saving part very quickly :)

@Fuckingdope ah colab, ok

i know xformers are/were problematic in general to install, the tutorials from nerdy show a specific version of xformers that worked at that time

have you tried installing the latest one? or those that are installed with a1111?

not sure why i have not seen it sooner, if you still have it, you can link on google drive and i can put it as a smaller version :)

@sevenof9247 thanks! uploaded the smaller file, i hope you like the new sample image :-)