usage: gen_single_image_diffusion.py --model_in_file MODEL_IN_FILE --img_in IMG_IN --dir_out DIR_OUT [--name NAME] [--img_width IMG_WIDTH] [--img_height IMG_HEIGHT] [--cpu] [--gpuid GPUID]
[--crop_width CROP_WIDTH] [--crop_height CROP_HEIGHT] [--bbox_width_factor BBOX_WIDTH_FACTOR] [--bbox_height_factor BBOX_HEIGHT_FACTOR]
[--sampling_steps SAMPLING_STEPS] [--seed SEED] [--mask_delta MASK_DELTA [MASK_DELTA ...]] [--mask_delta_ratio MASK_DELTA_RATIO [MASK_DELTA_RATIO ...]]
[--mask_square] [--sampling_method {ddpm,ddim}] [--cls_value CLS_VALUE] [--previous_frame PREVIOUS_FRAME] [--mask_in MASK_IN] [--ref-in REF_IN] [--bbox_in BBOX_IN]
[--nb_samples NB_SAMPLES] [--bbox_ref_id BBOX_REF_ID] [--cond_in COND_IN] [--cond_keep_ratio] [--cond_rotation COND_ROTATION]
[--cond_persp_horizontal COND_PERSP_HORIZONTAL] [--cond_persp_vertical COND_PERSP_VERTICAL] [--min_crop_bbox_ratio MIN_CROP_BBOX_RATIO]
[--alg_diffusion_cond_image_creation {y_t,previous_frame,sketch,canny,depth,hed,hough,low_res,sam,pix2pix}]
[--alg_diffusion_guidance_scale ALG_DIFFUSION_GUIDANCE_SCALE]
[--alg_diffusion_sketch_canny_thresholds ALG_DIFFUSION_SKETCH_CANNY_THRESHOLDS [ALG_DIFFUSION_SKETCH_CANNY_THRESHOLDS ...]]
[--alg_diffusion_super_resolution_downsample] [--alg_diffusion_sam_use_gaussian_filter] [--alg_diffusion_sam_no_sobel_filter]
[--alg_diffusion_sam_no_output_binary_sam] [--alg_diffusion_sam_redundancy_threshold ALG_DIFFUSION_SAM_REDUNDANCY_THRESHOLD]
[--alg_diffusion_sam_sobel_threshold ALG_DIFFUSION_SAM_SOBEL_THRESHOLD] [--alg_diffusion_sam_final_canny]
[--alg_diffusion_sam_min_mask_area ALG_DIFFUSION_SAM_MIN_MASK_AREA] [--alg_diffusion_sam_max_mask_area ALG_DIFFUSION_SAM_MAX_MASK_AREA]
[--alg_diffusion_sam_points_per_side ALG_DIFFUSION_SAM_POINTS_PER_SIDE] [--alg_diffusion_sam_no_sample_points_in_ellipse]
[--alg_diffusion_sam_crop_delta ALG_DIFFUSION_SAM_CROP_DELTA] [--model_prior_321_backwardcompatibility] [--alg_palette_ddim_num_steps ALG_PALETTE_DDIM_NUM_STEPS]
[--alg_palette_ddim_eta ALG_PALETTE_DDIM_ETA] [--f_s_weight_sam F_S_WEIGHT_SAM] [--data_refined_mask]
optional arguments:
--model_in_file MODEL_IN_FILE
file path to generator model (.pth file) (default: None)
--img_in IMG_IN image to transform (default: None)
--dir_out DIR_OUT The directory where to output result images (default: None)
--name NAME generated img name (default: img)
--img_width IMG_WIDTH
image width, defaults to model crop size (default: -1)
--img_height IMG_HEIGHT
image height, defaults to model crop size (default: -1)
--cpu whether to use CPU (default: False)
--gpuid GPUID which GPU to use (default: 0)
--crop_width CROP_WIDTH
crop width added on each side of the bbox (optional) (default: -1)
--crop_height CROP_HEIGHT
crop height added on each side of the bbox (optional) (default: -1)
--bbox_width_factor BBOX_WIDTH_FACTOR
bbox width added factor of original width (default: 0.0)
--bbox_height_factor BBOX_HEIGHT_FACTOR
bbox height added factor of original height (default: 0.0)
--sampling_steps SAMPLING_STEPS
number of sampling steps (default: -1)
--seed SEED random seed for reproducibility (default: -1)
--mask_delta MASK_DELTA [MASK_DELTA ...]
mask offset to allow generation of a bigger object, format : width (x) height (y) for each class or only one size if square (default: [[0]])
--mask_delta_ratio MASK_DELTA_RATIO [MASK_DELTA_RATIO ...]
ratio mask offset to allow generation of a bigger object, format : width (x),height (y) for each class or only one size if square (default: [[0]])
--mask_square whether to use square mask (default: False)
--sampling_method {ddpm,ddim}
choose the sampling method between ddpm and ddim (default: ddpm)
--cls_value CLS_VALUE
override input bbox classe for generation (default: -1)
--previous_frame PREVIOUS_FRAME
image to transform (default: None)
--mask_in MASK_IN mask used for image transformation (default: None)
--ref-in REF_IN image used as reference (default: None)
--bbox_in BBOX_IN bbox file used for masking (default: None)
--nb_samples NB_SAMPLES
nb of samples generated (default: 1)
--bbox_ref_id BBOX_REF_ID
bbox id to use (default: -1)
--cond_in COND_IN conditionning image to use (default: None)
--cond_keep_ratio
--cond_rotation COND_ROTATION
--cond_persp_horizontal COND_PERSP_HORIZONTAL
--cond_persp_vertical COND_PERSP_VERTICAL
--min_crop_bbox_ratio MIN_CROP_BBOX_RATIO
minimum crop/bbox ratio, allows to add context when bbox is larger than crop (default: None)
--alg_diffusion_cond_image_creation {y_t,previous_frame,sketch,canny,depth,hed,hough,low_res,sam,pix2pix}
how cond_image is created (default: None)
--alg_diffusion_guidance_scale ALG_DIFFUSION_GUIDANCE_SCALE
scale for classifier-free guidance, default is conditional DDPM only (default: 0.0)
--alg_diffusion_sketch_canny_thresholds ALG_DIFFUSION_SKETCH_CANNY_THRESHOLDS [ALG_DIFFUSION_SKETCH_CANNY_THRESHOLDS ...]
Canny thresholds (default: [0, 765])
--alg_diffusion_super_resolution_downsample
whether to downsample the image for super resolution (default: False)
--alg_diffusion_sam_use_gaussian_filter
whether to apply a gaussian blur to each SAM masks (default: False)
--alg_diffusion_sam_no_sobel_filter
whether to not use a Sobel filter on each SAM masks (default: True)
--alg_diffusion_sam_no_output_binary_sam
whether to not output binary sketch before Canny (default: True)
--alg_diffusion_sam_redundancy_threshold ALG_DIFFUSION_SAM_REDUNDANCY_THRESHOLD
redundancy threshold above which redundant masks are not kept (default: 0.62)
--alg_diffusion_sam_sobel_threshold ALG_DIFFUSION_SAM_SOBEL_THRESHOLD
sobel threshold in % of gradient magnitude (default: 0.7)
--alg_diffusion_sam_final_canny
whether to perform a Canny edge detection on sam sketch to soften the edges (default: False)
--alg_diffusion_sam_min_mask_area ALG_DIFFUSION_SAM_MIN_MASK_AREA
minimum area in proportion of image size for a mask to be kept (default: 0.001)
--alg_diffusion_sam_max_mask_area ALG_DIFFUSION_SAM_MAX_MASK_AREA
maximum area in proportion of image size for a mask to be kept (default: 0.99)
--alg_diffusion_sam_points_per_side ALG_DIFFUSION_SAM_POINTS_PER_SIDE
number of points per side of image to prompt SAM with (# of prompted points will be points_per_side**2) (default: 16)
--alg_diffusion_sam_no_sample_points_in_ellipse
whether to not sample the points inside an ellipse to avoid the corners of the image (default: True)
--alg_diffusion_sam_crop_delta ALG_DIFFUSION_SAM_CROP_DELTA
extend crop's width and height by 2*crop_delta before computing masks (default: True)
--model_prior_321_backwardcompatibility
whether to load models from previous version of JG. (default: False)
--alg_palette_ddim_num_steps ALG_PALETTE_DDIM_NUM_STEPS
number of steps for ddim sampling method (default: 10)
--alg_palette_ddim_eta ALG_PALETTE_DDIM_ETA
eta parameter for ddim variance (default: 0.5)
--f_s_weight_sam F_S_WEIGHT_SAM
path to sam weight for f_s, e.g. models/configs/sam/pretrain/sam_vit_b_01ec64.pth (default: models/configs/sam/pretrain/sam_vit_b_01ec64.pth)
--data_refined_mask whether to use refined mask with sam (default: False)
Bycob
commented
8 months ago