import torch.cuda import argparse from SUPIR.util import create_SUPIR_model, PIL2Tensor, Tensor2PIL, convert_dtype from PIL import Image from llava.llava_agent import LLavaAgent from CKPT_PTH import LLAVA_MODEL_PATH import os from torch.nn.functional import interpolate if torch.cuda.device_count() >= 2: SUPIR_device = 'cuda:0' LLaVA_device = 'cuda:1' elif torch.cuda.device_count() == 1: SUPIR_device = 'cuda:0' LLaVA_device = 'cuda:0' else: raise ValueError('Currently support CUDA only.') # hyparams here parser = argparse.ArgumentParser() parser.add_argument("--img_dir", type=str) parser.add_argument("--save_dir", type=str) parser.add_argument("--upscale", type=int, default=1) parser.add_argument("--SUPIR_sign", type=str, default='Q', choices=['F', 'Q']) parser.add_argument("--seed", type=int, default=1234) parser.add_argument("--min_size", type=int, default=1024) parser.add_argument("--edm_steps", type=int, default=50) parser.add_argument("--s_stage1", type=int, default=-1) parser.add_argument("--s_churn", type=int, default=5) parser.add_argument("--s_noise", type=float, default=1.003) parser.add_argument("--s_cfg", type=float, default=7.5) parser.add_argument("--s_stage2", type=float, default=1.) parser.add_argument("--num_samples", type=int, default=1) parser.add_argument("--a_prompt", type=str, default='Cinematic, High Contrast, highly detailed, taken using a Canon EOS R ' 'camera, hyper detailed photo - realistic maximum detail, 32k, Color ' 'Grading, ultra HD, extreme meticulous detailing, skin pore detailing, ' 'hyper sharpness, perfect without deformations.') parser.add_argument("--n_prompt", type=str, default='painting, oil painting, illustration, drawing, art, sketch, oil painting, ' 'cartoon, CG Style, 3D render, unreal engine, blurring, dirty, messy, ' 'worst quality, low quality, frames, watermark, signature, jpeg artifacts, ' 'deformed, lowres, over-smooth') parser.add_argument("--color_fix_type", type=str, default='Wavelet', choices=["None", "AdaIn", "Wavelet"]) parser.add_argument("--linear_CFG", action='store_true', default=True) parser.add_argument("--linear_s_stage2", action='store_true', default=False) parser.add_argument("--spt_linear_CFG", type=float, default=4.0) parser.add_argument("--spt_linear_s_stage2", type=float, default=0.) parser.add_argument("--ae_dtype", type=str, default="bf16", choices=['fp32', 'bf16']) parser.add_argument("--diff_dtype", type=str, default="fp16", choices=['fp32', 'fp16', 'bf16']) parser.add_argument("--no_llava", action='store_true', default=False) parser.add_argument("--loading_half_params", action='store_true', default=False) parser.add_argument("--use_tile_vae", action='store_true', default=False) parser.add_argument("--encoder_tile_size", type=int, default=512) parser.add_argument("--decoder_tile_size", type=int, default=64) parser.add_argument("--load_8bit_llava", action='store_true', default=False) args = parser.parse_args() print(args) use_llava = not args.no_llava # load SUPIR model = create_SUPIR_model('options/SUPIR_v0.yaml', SUPIR_sign=args.SUPIR_sign) if args.loading_half_params: model = model.half() if args.use_tile_vae: model.init_tile_vae(encoder_tile_size=args.encoder_tile_size, decoder_tile_size=args.decoder_tile_size) model.ae_dtype = convert_dtype(args.ae_dtype) model.model.dtype = convert_dtype(args.diff_dtype) model = model.to(SUPIR_device) # load LLaVA if use_llava: llava_agent = LLavaAgent(LLAVA_MODEL_PATH, device=LLaVA_device, load_8bit=args.load_8bit_llava, load_4bit=False) else: llava_agent = None os.makedirs(args.save_dir, exist_ok=True) for img_pth in os.listdir(args.img_dir): img_name = os.path.splitext(img_pth)[0] LQ_ips = Image.open(os.path.join(args.img_dir, img_pth)) LQ_img, h0, w0 = PIL2Tensor(LQ_ips, upsacle=args.upscale, min_size=args.min_size) LQ_img = LQ_img.unsqueeze(0).to(SUPIR_device)[:, :3, :, :] # step 1: Pre-denoise for LLaVA, resize to 512 LQ_img_512, h1, w1 = PIL2Tensor(LQ_ips, upsacle=args.upscale, min_size=args.min_size, fix_resize=512) LQ_img_512 = LQ_img_512.unsqueeze(0).to(SUPIR_device)[:, :3, :, :] clean_imgs = model.batchify_denoise(LQ_img_512) clean_PIL_img = Tensor2PIL(clean_imgs[0], h1, w1) # step 2: LLaVA if use_llava: captions = llava_agent.gen_image_caption([clean_PIL_img]) else: captions = [''] print(captions) # # step 3: Diffusion Process samples = model.batchify_sample(LQ_img, captions, num_steps=args.edm_steps, restoration_scale=args.s_stage1, s_churn=args.s_churn, s_noise=args.s_noise, cfg_scale=args.s_cfg, control_scale=args.s_stage2, seed=args.seed, num_samples=args.num_samples, p_p=args.a_prompt, n_p=args.n_prompt, color_fix_type=args.color_fix_type, use_linear_CFG=args.linear_CFG, use_linear_control_scale=args.linear_s_stage2, cfg_scale_start=args.spt_linear_CFG, control_scale_start=args.spt_linear_s_stage2) # save for _i, sample in enumerate(samples): Tensor2PIL(sample, h0, w0).save(f'{args.save_dir}/{img_name}_{_i}.png')