config.py

# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
# SPDX-License-Identifier: Apache-2.0

from dataclasses import dataclass
import torch

from src.eunms import Model_Type, Scheduler_Type

@dataclass
class RunConfig:
    model_type : Model_Type = Model_Type.SDXL_Turbo
    scheduler_type : Scheduler_Type = Scheduler_Type.EULER
    seed: int = 7865
    num_inference_steps: int = 4
    num_inversion_steps: int = 4
    inv_guidance_scale: float = 1.5
    guidance_scale: float = 5.0#2.0
    style_guidance_scale: float = 0.0#1.0
    content_guidance_scale: float = 0.0#0.5
    inv_style_guidance_scale: float = 0.0 # 5000
    inv_content_guidance_scale: float = 0.0
    inv_neg_style_guidance_scale: float = 0.0 # 5000
    inv_neg_content_guidance_scale: float = 0.0
    get_grad_guidance: bool = True
    inv_guidance: float = 0.9
    num_renoise_steps: int = 9
    max_num_renoise_steps_first_step: int = 5
    inversion_max_step: float = 1.0
    device = 'cuda'# if torch.cuda.is_available() else 'cpu'
    dtype = torch.float16
    useIP = True
    resolution: int = 1024

    # Average Parameters
    average_latent_estimations: bool = True
    average_first_step_range: tuple = (0, 5)
    average_step_range: tuple = (8, 10)

    # Noise Regularization
    noise_regularization_lambda_ac: float = 20.0
    noise_regularization_lambda_kl: float = 0.065
    noise_regularization_num_reg_steps: int = 4
    noise_regularization_num_ac_rolls: int = 5  

    # Noise Correction
    perform_noise_correction: bool = True

    # model path
    tile_controlnet_path = "./checkpoints/controlnet-tile-sdxl-1.0"
    canny_controlnet_path = "./checkpoints/MistoLine"
    depth_controlnet_path = "diffusers/controlnet-depth-sdxl-1.0-small"

    canny_controlnet_path_sd15 = './checkpoints/control_sd15_canny.pth'
    depth_controlnet_path_sd15 = './checkpoints/control_sd15_depth.pth'

    IP_path = "./checkpoints/IP-Adapter"
    clip_model_path = "./CSD_Score/models/ViT-L-14.pt"
    clip_path = "./CSD_Score/models/checkpoint.pth"
    base_model_path_sd15 =  "runwayml/stable-diffusion-v1-5"
    base_model_path = "stabilityai/stable-diffusion-xl-base-1.0"
    control_type = "tile_canny" # "tile" or "canny" or "depth" or "combine" or "tile_canny", where combine means use depth and canny
    choose_pipeline = ''#'sd15'

    # image param
    result_path = "./results"
    style_image_dir = "path/to/your/style/image/file" # such as ./data/style/7.jpg
    content_image_dir = "path/to/your/style/image/content/file" # such as ./data/content/7.jpg
    style_image_prompt = None # the prompt for style image, if None, use blip2 to generate
    resolution: int = 1024

    def __post_init__(self):
        pass