Skip to content

Latest commit

 

History

History
161 lines (139 loc) · 6.63 KB

DATASET_GENERATION.md

File metadata and controls

161 lines (139 loc) · 6.63 KB

Dataset Generation

Prerequisites

Installing Requirements

Run:

pip3 install -r requirements.txt

Downloading Models

Run:

python3 ./download_models.py

Processing Models

Run:

python3 ./process_models.py

Running Stable Diffusion scripts

Text To Image

Takes in a prompt and generates images

These are the available CLI arguments:

options:
  --prompt        PROMPT
                        An array of strings to help guide the generation process
  --batch_size   BATCH_SIZE
                        How many images to generate at once
  --output        OUTPUT
                        Number of folders to
  --sampler       SAMPLER
                        Name of the sampler to use
  --checkpoint_path     CHECKPOINT_PATH
                        Path to the checkpoint file
  --flash         FLASH
                        Whether to use flash attention
  --steps         STEPS
                        Number of steps to use
  --cgf_scale         SCALE
                        Unconditional guidance scale: eps = eps(x, empty) + scale * (eps(x, cond) - eps(x, empty))
  --seed      SEED
                        Array of seed for the image generation: example '0, 1, 0, 7', Its better if the size of the array is the same as the number of generated images
  --low-vram      LOW_VRAM
                        Limit vram usage
  --force_cpu     FORCE_CPU
                        Force cpu usage
  --cuda_device   CUDA_DEVICE
                        Cuda device to use for generation process
  --num_images    NUM_IMAGES
                        Number of images to output

Example Usage:

python3 ./scripts/text_to_image.py --prompt "character, chibi, waifu, side scrolling, white background, centered" --checkpoint_path "./input/model/v1-5-pruned-emaonly.safetensors" --batch_size 1 --num_images 6

Random Prompts Generation and Disturbing Embeddings Image Generation

Try running:

python3 ./scripts/data_bounding_box_and_score_and_embedding_dataset.py --num_iterations 10
  • cfg_strength: Configuration strength. Defaults to 12.
  • embedded_prompts_dir: The path to the directory containing the embedded prompts tensors. Defaults to a constant EMBEDDED_PROMPTS_DIR, which is expected to be './input/embedded_prompts/'
  • num_iterations: The number of iterations to batch-generate images. Defaults to 8.
  • batch_size: The number of images to generate per batch. Defaults to 1.
  • seed: The noise seed used to generate the images. Defaults to random int `0 to 2^24``.
  • noise_multiplier: The multiplier for the amount of noise used to disturb the prompt embedding. Defaults to 0.008.
  • cuda_device: The CUDA device to use. Defaults to 'get_device()'.
  • clear_output_dir: If True, the output directory will be cleared before generating images. Defaults to False.

Generate Images From Prompt Generator

To generate images from random prompt, these are the available CLI arguments:

options:
  -h, --help            show this help message and exit
  --prompts_file PROMPTS_FILE
                        Path to the file containing the prompts, each on a line (default: './input/prompts.txt')
  --batch_size BATCH_SIZE
                        How many images to generate at once (default: 1)
  --output OUTPUT       Path to the output directory (default: /output)
  --sampler SAMPLER     Name of the sampler to use (default: ddim)
  --checkpoint_path CHECKPOINT_PATH
                        Path to the checkpoint file (default: './input/model/v1-5-pruned-emaonly.safetensors')
  --flash               whether to use flash attention
  --steps STEPS         Number of steps to use (default: 20)
  --cfg_scale CFG_SCALE
                        unconditional guidance scale: eps = eps(x, empty) + scale * (eps(x, cond) - eps(x, empty))
  --low_vram            limit VRAM usage
  --force_cpu           force CPU usage
  --cuda_device CUDA_DEVICE
                        cuda device to use for generation
  --num_images NUM_IMAGES
                        How many images to generate (default: 1)
  --seed SEED           Seed for the image generation (default: )
  --output_metadata     outputs the metadata
  --image_width IMAGE_WIDTH
                        Generate image width (default: 512)
  --image_height IMAGE_HEIGHT
                        Generate image height (default: 512)
  --num_datasets NUM_DATASETS
                        Number of datasets to generate (default: 1)
  --num_phrases [NUM_PHRASES]
                        The number of phrases for the prompt to generate

python3 ./scripts/generate_images_from_prompt_generator.py --checkpoint_path "./input/model/v1-5-pruned-emaonly.safetensors" --cfg_scale 7 --num_images 10 --num_phrases 12 --output "./output/"

Generate Images From Prompt List Dataset Embeddings

To generate images from prompt list dataset embeddings, these are the available CLI arguments:

options:
  -h, --help            show this help message and exit
  --batch_size BATCH_SIZE
                        How many images to generate at once (default: 1)
  --output OUTPUT       Path to the output directory (default: ./output)
  --sampler SAMPLER     Name of the sampler to use (default: ddim)
  --checkpoint_path CHECKPOINT_PATH
                        Path to the checkpoint file (default: './input/model/v1-5-pruned-emaonly.safetensors')
  --flash               whether to use flash attention
  --steps STEPS         Number of steps to use (default: 20)
  --cfg_scale CFG_SCALE
                        unconditional guidance scale: eps = eps(x, empty) + scale * (eps(x, cond) - eps(x, empty))
  --force_cpu           force CPU usage
  --cuda_device CUDA_DEVICE
                        cuda device to use for generation
  --num_images NUM_IMAGES
                        How many images to generate (default: 1)
  --seed SEED           Seed for the image generation (default: )
  --image_width IMAGE_WIDTH
                        Generate image width (default: 512)
  --image_height IMAGE_HEIGHT
                        Generate image height (default: 512)
  --num_datasets NUM_DATASETS
                        Number of datasets to generate (default: 1)
  --image_batch_size IMAGE_BATCH_SIZE
                        Number of batches (default: 1)
  --prompt_list_dataset_path PROMPT_LIST_DATASET_PATH
                        The path to prompt list dataset zip

python3 ./scripts/generate_images_from_prompt_list_embeddings.py --checkpoint_path "./input/model/sd/v1-5-pruned-emaonly/v1-5-pruned-emaonly.safetensors" --cfg_scale 7 --num_images 2 --output ./output/generated-dataset-from-prompt-list --prompt_list_dataset_path ./test/test_zip_files/prompt_list_civitai_2_test.zip