Instance images root doesn't exists.

[AFMSB]

I have tried to run your code, the problem is i am getting this error, i am using the same dataset i used with the ai-toolkit tool:

Starting 1 job(s)
Starting job train_lora_sdxl_24gb_1.0.yaml
Using device: cuda:0
Using TF32 for faster training on Ampere GPUs
Using matmul precision: high
Using cuDNN benchmark
Using seed: 42
Using variant: fp16
Using 4 workers for data loading
Walking seed
Loading noise scheduler
{'variance_type', 'clip_sample_range', 'thresholding', 'rescale_betas_zero_snr', 'dynamic_thresholding_ratio'} was not found in config. Values will be initialized to default values.
Using DDPMScheduler
Loading text encoder 1
You are using a model of type clip_text_model to instantiate a model of type . This is not supported for all configurations of models and can yield errors.
Loading text encoder 2
You are using a model of type clip_text_model to instantiate a model of type . This is not supported for all configurations of models and can yield errors.
Loading VAE
{'latents_mean', 'use_quant_conv', 'use_post_quant_conv', 'mid_block_add_attention', 'shift_factor', 'latents_std'} was not found in config. Values will be initialized to default values.
Loading unet
{'dropout', 'attention_type', 'reverse_transformer_layers_per_block'} was not found in config. Values will be initialized to default values.
Adding LoRA adapters
Loading optimizer
Loading datasets
Job train_lora_sdxl_24gb_1.0.yaml failed with error: Instance images root doesn't exists.
Traceback (most recent call last):
File "/workspace/lorakit/.venv/lib/python3.10/site-packages/lorakit/cli/lorakit.py", line 31, in main
job.run()
File "/workspace/lorakit/.venv/lib/python3.10/site-packages/lorakit/train.py", line 762, in run
train_dataset = DreamBoothDataset(self._dataset_folder, self._instant_prompt, self._class_prompt, resolution=self._resolution)
File "/workspace/lorakit/.venv/lib/python3.10/site-packages/lorakit/datasets.py", line 41, in init
raise ValueError("Instance images root doesn't exists.")
ValueError: Instance images root doesn't exists.

[omidsakhi]

The condition that causes this error is very simple. It checks to see if folder exists or not. Maybe there is a typo in the path?

[AFMSB]

@omidsakhi i have already fixed it, but can you provide examples of what to put in these fields:

instant_prompt: "A highly detailed, photorealistic portrait of andre" # prompt to use for the instant prompt
class_prompt: "man" # prompt to use for the class prompt

prompts:
- "Create a photorealistic professional portrait of [trigger] wearing a sharp, dark business suit and tie. The background is a clean, solid color [trigger] is smiling confidently, with a direct gaze that conveys professionalism and authority. The lighting is soft and natural, highlighting their features. Perfect for a corporate headshot."

neg:
- "black and white, monochrome, grayscale, sepia, disfigured face, extra fingers, missing limbs, cartoonish, low quality, blurry, text, watermark"

[omidsakhi]

instant_prompt is the trigger word. lorakit does not follow the implementation of trigger word like ai-toolkit yet. In this case it follows the implementation of HuggingFace Auto-Train Advanced. So here you should write "SKS" or "TOK" or something unique. It can also be "Photo of SKS" or "Photo of TOK". Do not put anyhting such as high-quality and stuff because they will be assigned to you training images and usually they are not high-quality.

class_prompt is what the SKS or TOK is supposed to be. For example man or woman.

prompts: is an array of positive prompts for sampling during the training. So you can see how the training is progressing epoch to epoch.
neg: is an array of negative prompts that is supplied with positive prompts. The array sizes must match.

[AFMSB]

ok, but shall i place "TOK" in my prompts?, Do you have some examples for the negative ones?

[omidsakhi]

Yes. Sorry I forgot. Please put TOK in the prompts until we implement a mechanism similar to [trigger] later.

[AFMSB]

also did you manage to accomplish the realistic effect of flux, i have been trying everything but didnt reach it, maybe its my config, i always get plastic skin

[AFMSB]

@omidsakhi adter i changed my initial prompt like you said my loras dont even look like the person on the dataset

[omidsakhi]

Would you please share your config file?