Datasets:

stablellama
/

Krea-2-Raw_samples

license: cc-by-4.0
task_categories:
  - text-to-image
language:
  - en
tags:
  - LoRA
  - LyCROIS
  - LoKR
  - training
  - finetuning
pretty_name: Krea 2 Raw samples
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: test
        path:
          - data/metadata.csv
          - data/**/*.jxl

This dataset is a highly diverse set of high quality images generated with Krea 2 Raw.

NOTE: Raw is not intended for image generation, so do not use these images to judge the quality of the model. Raw is intended for training, as are the samples in this dataset as they can be used for regularization.

Possible uses

Regularization images for training models based on Krea 2 Raw
Quality testing

Data source

The images were created in ComfyUI with the bf16 version of Krea 2 Raw. For each prompt were four images generated, all are (without any cherry picking) included in the corresponding dataset directories.

bf16 - published full model weights
1024x1024 pixels - native resolution for 1:1 aspect ratio images
batch size = 4
fixed seed (seed = 1)
workflow is the ComfyUI default for Krea 2
52 steps (as used by Krea at their Huggingface model documentation)
CGF = 3.5 (as used by Krea)
sampler: Euler
scheduler: simple

Data content

The data attempts to cover a very wide range of concepts. They are categorized into:

animals & creatures (Biological Diversity) (4x 80 images)
architecture (4x 60 images)
clothing (4x 70 images)
composition (4x 40 images)
DiffusionDB, originals ones and prompts but refined by a LLM (4x 170 images)
dynamic poses (4x 80 images)
hands (4x 80 images)
handicraft (4x 20 images)
humans (4x 80 images)
interaction (4x 40 images)
landscape (4x 20 images)
logo, English (4x 80 images)
noteable persons (4x 17 images)
objects (4x 100 images)
style (4x 80 images)
Chinese text (4x 100 images)
English text (4x 100 images)

In total are 1217 different prompts in the dataset with 4 images each, resulting in 4x 1217 = 4868 images.
Contact prints for each category are at https://huggingface.co/datasets/stablellama/Krea-2-Raw_samples/tree/main/contact_print

The images are stored in lossless compressed JPEG XL format.

NOTE: Krea 2 Raw does tend to create good as well as bad images (e.g. buildings with bad lines). This can easily be seen in the provided images where each batch of 4 images is uploaded without any cherry picking. This implies, that when you want to use these images as regularization data for LoRA training, you should filter the images you are using. You can also use the stablellama/Krea-2-Raw_samples_Best_of dataset, which is already manually filtered and contains only one image per prompt.