-
Notifications
You must be signed in to change notification settings - Fork 0
/
random_walk.yaml
42 lines (36 loc) · 3.25 KB
/
random_walk.yaml
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
# STABLE DIFFUSION EXPERIMENT CONFIGURATION
# This config file contains the experiment configuration for the Stable Diffusion txt2img model from the diffusers library.
# IDENTIFIER CONFIGURATIONS
# Specifies the model and experiment identifiers.
# DO NOT CHANGE THESE!
model_identifier: txt2img
exp_identifier: random-walk
# MODEL CONFIGURATIONS
# Specifies the model configurations.
model_id: CompVis/stable-diffusion-v1-4 # Name of the Stable Diffusion repository on HuggingFace (e.g. stabilityai/stable-diffusion-2-1) or the path to the cloned repository (e.g. /mypath/stable-diffusion-2-1).
scheduler: DPMSolverMultistepScheduler # Name of the scheduler algorithm.
att_slicing: True # Whether attention slicing should be used (reduces memory consumption during the diffusion process at the cost of speed).
vae_slicing: True # Whether VAE slicing should be used (reduces memory consumption during the decoding stage at the cost of speed).
vae_tiling: False # Whether VAE tiling should be used (reduces memory consumption during the decoding stage at the cost of speed).
enable_xformers: False # Whether to enable xFormers for optimized performance in the attention blocks (requires the xformers package).
gpu_id: 0 # GPU index.
diffusion_steps: 50 # Amount of diffusion steps to perform (higher values increase quality at the cost of speed).
guidance_scale: 9.5 # Guidance scale factor for classifier free guidance (higher values lead to better correspondence to the prompt, while lower values increase diversity).
# EXPERIMENT CONFIGURATIONS
# Specifies the experiment configurations.
output_path: ./experiments # Path for storing the experiment results (a new folder will be placed at the specified location).
gif_frame_dur: 75 # Specifies the frame duration in milliseconds for the produced gifs.
prompt_rand_walk: True # Whether to perform a random walk within the latent space of the text encoder.
noise_rand_walk: True # Whether to perform a random walk within the latent space of the VAE.
walk_steps: 40 # Specifies the amount of steps for the random walk.
step_size: 0.0095 # Specifies the step size for each step of the random walk.
walk_directions: 2 # Specifies the amount of latent space directions to be explored by the random walk.
# PROMPT CONFIGURATION
prompt: king walking on water, fantasy art|low resolution # Input prompt where the positive part is separated from the negative part by a vertical line "|" without any whitespace in between.
load_prompt_embeds: None # Path to a local file containing the prompt embeddings. Caution the parameter "prompt" does not apply, if a pre-generated prompt embedding is loaded from a file.
# LATENT NOISE CONFIGURATION
rand_seed: 0 # Random seed for sampling reproducible latent noise.
height: 256 # Image height of the desired VAE output (used for computing the latent noise height).
width: 256 # Image width of the desired VAE output (used for computing the latent noise width).
images_per_prompt: 1 # Amount of images to generate per prompt (specifies the batch dimension of the latent noise).
load_latent_noise: None # Path to a local file containing the latent noise tensor. Caution the parameters "rand_seed", "height", "width" and "images_per_prompt" do not apply, if a pre-generated latent noise tensor is loaded from a file.