app.py

# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
# //
# // Licensed under the Apache License, Version 2.0 (the "License");
# // you may not use this file except in compliance with the License.
# // You may not obtain a copy of the License at
# //
# //     http://www.apache.org/licenses/LICENSE-2.0
# //
# // Unless required by applicable law or agreed to in writing, software
# // distributed under the License is distributed on an "AS IS" BASIS,
# // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# // See the License for the specific language governing permissions and
# // limitations under the License.
import spaces
import subprocess
import os
import sys

# --- ETAPA 1: Clonar o Repositório Oficial do GitHub ---
repo_name = "SeedVR"
if not os.path.exists(repo_name):
    print(f"Clonando o repositório {repo_name} do GitHub...")
    subprocess.run(f"git clone https://github.com/ByteDance-Seed/{repo_name}.git", shell=True, check=True)

# --- ETAPA 2: Mudar para o Diretório e Configurar o Ambiente ---
os.chdir(repo_name)
print(f"Diretório de trabalho alterado para: {os.getcwd()}")

# Adicionar o diretório ao path do Python para que as importações funcionem
sys.path.insert(0, os.path.abspath('.'))
print(f"Diretório atual adicionado ao sys.path.")

# --- ETAPA 3: Instalar Dependências Conforme as Instruções ---
python_executable = sys.executable
print("Instalando dependências do requirements.txt...")
subprocess.run([python_executable, "-m", "pip", "install", "-r", "requirements.txt"], check=True)

print("Instalando flash-attn...")
subprocess.run([python_executable, "-m", "pip", "install", "flash-attn==2.5.9.post1", "--no-build-isolation"], check=True)

from pathlib import Path
from urllib.parse import urlparse
from torch.hub import download_url_to_file, get_dir

# Função auxiliar para downloads
def load_file_from_url(url, model_dir='.', progress=True, file_name=None):
    os.makedirs(model_dir, exist_ok=True)
    if not file_name:
        parts = urlparse(url)
        file_name = os.path.basename(parts.path)
    cached_file = os.path.join(model_dir, file_name)
    if not os.path.exists(cached_file):
        print(f'Baixando: "{url}" para {cached_file}\n')
        download_url_to_file(url, cached_file, hash_prefix=None, progress=progress)
    return cached_file

# Baixar e instalar Apex pré-compilado (crucial para o ambiente do Spaces)
apex_url = 'https://huggingface.co/ByteDance-Seed/SeedVR2-3B/resolve/main/apex-0.1-cp310-cp310-linux_x86_64.whl'
apex_wheel_path = load_file_from_url(url=apex_url)
print("Instalando Apex a partir do wheel baixado...")
subprocess.run([python_executable, "-m", "pip", "install", "--force-reinstall", "--no-cache-dir", apex_wheel_path], check=True)
print("✅ Configuração do Apex concluída.")

# --- ETAPA 4: Baixar os Modelos Pré-treinados ---
print("Baixando modelos pré-treinados...")
pretrain_model_url = {
    'vae': 'https://huggingface.co/ByteDance-Seed/SeedVR2-3B/resolve/main/ema_vae.pth',
    'dit': 'https://huggingface.co/ByteDance-Seed/SeedVR2-3B/resolve/main/seedvr2_ema_3b.pth',
    'pos_emb': 'https://huggingface.co/ByteDance-Seed/SeedVR2-3B/resolve/main/pos_emb.pt',
    'neg_emb': 'https://huggingface.co/ByteDance-Seed/SeedVR2-3B/resolve/main/neg_emb.pt',
}

Path('./ckpts').mkdir(exist_ok=True)
for key, url in pretrain_model_url.items():
    model_dir = './ckpts' if key in ['vae', 'dit'] else '.'
    load_file_from_url(url=url, model_dir=model_dir)

# Baixar vídeos de exemplo
torch.hub.download_url_to_file('https://huggingface.co/datasets/Iceclear/SeedVR_VideoDemos/resolve/main/seedvr_videos_crf23/aigc1k/23_1_lq.mp4', '01.mp4')
torch.hub.download_url_to_file('https://huggingface.co/datasets/Iceclear/SeedVR_VideoDemos/resolve/main/seedvr_videos_crf23/aigc1k/28_1_lq.mp4', '02.mp4')
torch.hub.download_url_to_file('https://huggingface.co/datasets/Iceclear/SeedVR_VideoDemos/resolve/main/seedvr_videos_crf23/aigc1k/2_1_lq.mp4', '03.mp4')
print("✅ Setup completo. Iniciando a aplicação...")

# --- ETAPA 5: Executar a Aplicação Principal ---
import torch
import mediapy
from einops import rearrange
from omegaconf import OmegaConf
import datetime
from tqdm import tqdm
import gc
from PIL import Image
import gradio as gr
import uuid
import mimetypes
import torchvision.transforms as T
from torchvision.transforms import Compose, Lambda, Normalize
from torchvision.io.video import read_video

from data.image.transforms.divisible_crop import DivisibleCrop
from data.image.transforms.na_resize import NaResize
from data.video.transforms.rearrange import Rearrange
from common.config import load_config
from common.distributed import init_torch
from common.distributed.advanced import init_sequence_parallel
from common.seed import set_seed
from common.partition import partition_by_size
from projects.video_diffusion_sr.infer import VideoDiffusionInfer
from common.distributed.ops import sync_data

os.environ["MASTER_ADDR"] = "127.0.0.1"
os.environ["MASTER_PORT"] = "12355"
os.environ["RANK"] = str(0)
os.environ["WORLD_SIZE"] = str(1)

if os.path.exists("projects/video_diffusion_sr/color_fix.py"):
    from projects.video_diffusion_sr.color_fix import wavelet_reconstruction
    use_colorfix = True
else:
    use_colorfix = False

def configure_runner():
    config = load_config('configs_3b/main.yaml')
    runner = VideoDiffusionInfer(config)
    OmegaConf.set_readonly(runner.config, False)
    init_torch(cudnn_benchmark=False, timeout=datetime.timedelta(seconds=3600))
    runner.configure_dit_model(device="cuda", checkpoint='ckpts/seedvr2_ema_3b.pth')
    runner.configure_vae_model()
    if hasattr(runner.vae, "set_memory_limit"):
        runner.vae.set_memory_limit(**runner.config.vae.memory_limit)
    return runner

def generation_step(runner, text_embeds_dict, cond_latents):
    def _move_to_cuda(x): return [i.to("cuda") for i in x]
    noises = [torch.randn_like(latent) for latent in cond_latents]
    aug_noises = [torch.randn_like(latent) for latent in cond_latents]
    noises, aug_noises, cond_latents = sync_data((noises, aug_noises, cond_latents), 0)
    noises, aug_noises, cond_latents = list(map(_move_to_cuda, (noises, aug_noises, cond_latents)))
    def _add_noise(x, aug_noise):
        t = torch.tensor([100.0], device="cuda")
        shape = torch.tensor(x.shape[1:], device="cuda")[None]
        t = runner.timestep_transform(t, shape)
        return runner.schedule.forward(x, aug_noise, t)
    conditions = [runner.get_condition(n, task="sr", latent_blur=_add_noise(l, an)) for n, an, l in zip(noises, aug_noises, cond_latents)]
    with torch.no_grad(), torch.autocast("cuda", torch.bfloat16, enabled=True):
        video_tensors = runner.inference(noises=noises, conditions=conditions, **text_embeds_dict)
    return [rearrange(v, "c t h w -> t c h w") for v in video_tensors]

@spaces.GPU
def generation_loop(video_path, seed=666, fps_out=24):
    if video_path is None: return None, None, None
    runner = configure_runner()
    text_embeds = {"texts_pos": [torch.load('pos_emb.pt').to("cuda")], "texts_neg": [torch.load('neg_emb.pt').to("cuda")]}
    runner.configure_diffusion()
    set_seed(int(seed))
    os.makedirs("output", exist_ok=True)
    video_transform = Compose([NaResize(1024), DivisibleCrop(16), Normalize(0.5, 0.5), Rearrange("t c h w -> c t h w")])
    media_type, _ = mimetypes.guess_type(video_path)
    is_video = media_type and media_type.startswith("video")
    if is_video:
        video, _, _ = read_video(video_path, output_format="TCHW")
        video = video[:121] / 255.0
        output_path = os.path.join("output", f"{uuid.uuid4()}.mp4")
    else:
        video = T.ToTensor()(Image.open(video_path).convert("RGB")).unsqueeze(0)
        output_path = os.path.join("output", f"{uuid.uuid4()}.png")
    cond_latents = [video_transform(video.to("cuda"))]
    ori_length = cond_latents[0].size(2)
    cond_latents = runner.vae_encode(cond_latents)
    samples = generation_step(runner, text_embeds, cond_latents)
    sample = samples[0][:ori_length].cpu()
    sample = rearrange(sample, "t c h w -> t h w c").clip(-1, 1).add(1).mul(127.5).byte().numpy()
    if is_video:
        mediapy.write_video(output_path, sample, fps=fps_out)
        return None, output_path, output_path
    else:
        mediapy.write_image(output_path, sample[0])
        return output_path, None, output_path

with gr.Blocks(title="SeedVR") as demo:
    gr.HTML(f"""<div style='text-align:center; margin-bottom: 10px;'><img src='file/{os.path.abspath("assets/seedvr_logo.png")}' style='height:40px;'/></div>...""")
    with gr.Row():
        input_file = gr.File(label="Carregar Imagem ou Vídeo")
        with gr.Column():
            seed = gr.Number(label="Seed", value=42)
            fps = gr.Number(label="FPS de Saída", value=24)
    run_button = gr.Button("Executar")
    output_image = gr.Image(label="Imagem de Saída")
    output_video = gr.Video(label="Vídeo de Saída")
    download_link = gr.File(label="Baixar Resultado")
    run_button.click(fn=generation_loop, inputs=[input_file, seed, fps], outputs=[output_image, output_video, download_link])
    gr.Examples(examples=[["01.mp4", 42, 24], ["02.mp4", 42, 24], ["03.mp4", 42, 24]], inputs=[input_file, seed, fps])
    gr.HTML("""<hr>...""")

demo.queue().launch(share=True)