Update app.py

app.py

@@ -1,111 +1,217 @@
-import os, sys
+import os
+import platform
+import uuid
+import shutil
+from pydub import AudioSegment
+import spaces
+import torch
 import gradio as gr
-from src.gradio_demo import SadTalker  
+from huggingface_hub import snapshot_download
 
+from examples.get_examples import get_examples
+from src.facerender.pirender_animate import AnimateFromCoeff_PIRender
+from src.utils.preprocess import CropAndExtract
+from src.test_audio2coeff import Audio2Coeff
+from src.facerender.animate import AnimateFromCoeff
+from src.generate_batch import get_data
+from src.generate_facerender_batch import get_facerender_data
+from src.utils.init_path import init_path
 
-try:
-    import webui  # in webui
-    in_webui = True
-except:
-    in_webui = False
+checkpoint_path = 'checkpoints'
+config_path = 'src/config'
+device = "cuda" if torch.cuda.is_available(
+) else "mps" if platform.system() == 'Darwin' else "cpu"
 
+os.environ['TORCH_HOME'] = checkpoint_path
+snapshot_download(repo_id='vinthony/SadTalker-V002rc',
+                  local_dir=checkpoint_path, local_dir_use_symlinks=True)
 
-def toggle_audio_file(choice):
-    if choice == False:
-        return gr.update(visible=True), gr.update(visible=False)
+
+def mp3_to_wav(mp3_filename, wav_filename, frame_rate):
+    AudioSegment.from_file(file=mp3_filename).set_frame_rate(
+        frame_rate).export(wav_filename, format="wav")
+
+
+@spaces.GPU(duration=120)
+def generate_video(source_image, driven_audio, preprocess='crop', still_mode=False, use_enhancer=False,
+                   batch_size=1, size=256, pose_style=0, facerender='facevid2vid', exp_scale=1.0,
+                   use_ref_video=False, ref_video=None, ref_info=None, use_idle_mode=False,
+                   length_of_audio=0, use_blink=True, result_dir='./results/'):
+    # Initialize models and paths
+    sadtalker_paths = init_path(
+        checkpoint_path, config_path, size, False, preprocess)
+    audio_to_coeff = Audio2Coeff(sadtalker_paths, device)
+    preprocess_model = CropAndExtract(sadtalker_paths, device)
+    animate_from_coeff = AnimateFromCoeff(sadtalker_paths, device) if facerender == 'facevid2vid' and device != 'mps' \
+        else AnimateFromCoeff_PIRender(sadtalker_paths, device)
+
+    # Create directories for saving results
+    time_tag = str(uuid.uuid4())
+    save_dir = os.path.join(result_dir, time_tag)
+    os.makedirs(save_dir, exist_ok=True)
+    input_dir = os.path.join(save_dir, 'input')
+    os.makedirs(input_dir, exist_ok=True)
+
+    # Process source image
+    pic_path = os.path.join(input_dir, os.path.basename(source_image))
+    shutil.move(source_image, input_dir)
+
+    # Process driven audio
+    if driven_audio and os.path.isfile(driven_audio):
+        audio_path = os.path.join(input_dir, os.path.basename(driven_audio))
+        if '.mp3' in audio_path:
+            mp3_to_wav(driven_audio, audio_path.replace('.mp3', '.wav'), 16000)
+            audio_path = audio_path.replace('.mp3', '.wav')
+        else:
+            shutil.move(driven_audio, input_dir)
+    elif use_idle_mode:
+        audio_path = os.path.join(
+            input_dir, 'idlemode_'+str(length_of_audio)+'.wav')
+        AudioSegment.silent(
+            duration=1000*length_of_audio).export(audio_path, format="wav")
     else:
-        return gr.update(visible=False), gr.update(visible=True)
-    
-def ref_video_fn(path_of_ref_video):
-    if path_of_ref_video is not None:
-        return gr.update(value=True)
+        assert use_ref_video and ref_info == 'all'
+
+    # Process reference video
+    if use_ref_video and ref_info == 'all':
+        ref_video_videoname = os.path.splitext(os.path.split(ref_video)[-1])[0]
+        audio_path = os.path.join(save_dir, ref_video_videoname+'.wav')
+        os.system(
+            f"ffmpeg -y -hide_banner -loglevel error -i {ref_video} {audio_path}")
+        ref_video_frame_dir = os.path.join(save_dir, ref_video_videoname)
+        os.makedirs(ref_video_frame_dir, exist_ok=True)
+        ref_video_coeff_path, _, _ = preprocess_model.generate(
+            ref_video, ref_video_frame_dir, preprocess, source_image_flag=False)
     else:
-        return gr.update(value=False)
-
-def sadtalker_demo(checkpoint_path='checkpoints', config_path='src/config', warpfn=None):
-
-    sad_talker = SadTalker(checkpoint_path, config_path, lazy_load=True)
-
-    with gr.Blocks(analytics_enabled=False) as sadtalker_interface:
-        gr.Markdown("<div align='center'> <h2> 😭 SadTalker: Learning Realistic 3D Motion Coefficients for Stylized Audio-Driven Single Image Talking Face Animation (CVPR 2023) </span> </h2> \
-                    <a style='font-size:18px;color: #efefef' href='https://arxiv.org/abs/2211.12194'>Arxiv</a> &nbsp;&nbsp;&nbsp;&nbsp;&nbsp; \
-                    <a style='font-size:18px;color: #efefef' href='https://sadtalker.github.io'>Homepage</a>  &nbsp;&nbsp;&nbsp;&nbsp;&nbsp; \
-                     <a style='font-size:18px;color: #efefef' href='https://github.com/Winfredy/SadTalker'> Github </div>")
-        
-        with gr.Row().style(equal_height=False):
-            with gr.Column(variant='panel'):
-                with gr.Tabs(elem_id="sadtalker_source_image"):
-                    with gr.TabItem('Upload image'):
+        ref_video_coeff_path = None
+
+    # Preprocess source image
+    first_frame_dir = os.path.join(save_dir, 'first_frame_dir')
+    os.makedirs(first_frame_dir, exist_ok=True)
+    first_coeff_path, crop_pic_path, crop_info = preprocess_model.generate(
+        pic_path, first_frame_dir, preprocess, True, size)
+    if first_coeff_path is None:
+        raise AttributeError("No face is detected")
+
+    # Determine reference coefficients
+    ref_pose_coeff_path, ref_eyeblink_coeff_path = None, None
+    if use_ref_video:
+        if ref_info == 'pose':
+            ref_pose_coeff_path = ref_video_coeff_path
+        elif ref_info == 'blink':
+            ref_eyeblink_coeff_path = ref_video_coeff_path
+        elif ref_info == 'pose+blink':
+            ref_pose_coeff_path = ref_eyeblink_coeff_path = ref_video_coeff_path
+    else:
+        ref_pose_coeff_path = ref_eyeblink_coeff_path = None
+
+    # Generate coefficients from audio or reference video
+    if use_ref_video and ref_info == 'all':
+        coeff_path = ref_video_coeff_path
+    else:
+        batch = get_data(first_coeff_path, audio_path, device, ref_eyeblink_coeff_path=ref_eyeblink_coeff_path,
+                         still=still_mode, idlemode=use_idle_mode, length_of_audio=length_of_audio, use_blink=use_blink)
+        coeff_path = audio_to_coeff.generate(
+            batch, save_dir, pose_style, ref_pose_coeff_path)
+
+    # Generate video from coefficients
+    data = get_facerender_data(coeff_path, crop_pic_path, first_coeff_path, audio_path, batch_size, still_mode=still_mode,
+                               preprocess=preprocess, size=size, expression_scale=exp_scale, facemodel=facerender)
+    return_path = animate_from_coeff.generate(data, save_dir, pic_path, crop_info, enhancer='gfpgan' if use_enhancer else None,
+                                              preprocess=preprocess, img_size=size)
+    video_name = data['video_name']
+    print(f'The generated video is named {video_name} in {save_dir}')
+
+    return return_path
+
+
+# Gradio UI
+with gr.Blocks(analytics_enabled=False) as demo:
+    with gr.Row():
+        with gr.Column(variant='panel'):
+            with gr.Tabs(elem_id="sadtalker_source_image"):
+                with gr.TabItem('Source image'):
+                    with gr.Row():
+                        source_image = gr.Image(
+                            label="Source image", sources="upload", type="filepath", elem_id="img2img_image")
+
+            with gr.Tabs(elem_id="sadtalker_driven_audio"):
+                with gr.TabItem('Driving Methods'):
+                    gr.Markdown(
+                        "Possible driving combinations: <br> 1. Audio only 2. Audio/IDLE Mode + Ref Video(pose, blink, pose+blink) 3. IDLE Mode only 4. Ref Video only (all) ")
+
+                    with gr.Row():
+                        driven_audio = gr.Audio(
+                            label="Input audio", sources="upload", type="filepath")
+                        driven_audio_no = gr.Audio(
+                            label="Use IDLE mode, no audio is required", sources="upload", type="filepath", visible=False)
+
+                        with gr.Column():
+                            use_idle_mode = gr.Checkbox(
+                                label="Use Idle Animation")
+                            length_of_audio = gr.Number(
+                                value=5, label="The length(seconds) of the generated video.")
+                            use_idle_mode.change(lambda choice: (gr.update(visible=not choice), gr.update(visible=choice)),
+                                                 inputs=use_idle_mode, outputs=[driven_audio, driven_audio_no])
+
+                    with gr.Row():
+                        ref_video = gr.Video(
+                            label="Reference Video", sources="upload", elem_id="vidref")
+
+                        with gr.Column():
+                            use_ref_video = gr.Checkbox(
+                                label="Use Reference Video")
+                            ref_info = gr.Radio(['pose', 'blink', 'pose+blink', 'all'], value='pose', label='Reference Video',
+                                                info="How to borrow from reference Video?((fully transfer, aka, video driving mode))")
+
+                        ref_video.change(lambda path: gr.update(
+                            value=path is not None), inputs=ref_video, outputs=use_ref_video)
+
+        with gr.Column(variant='panel'):
+            with gr.Tabs(elem_id="sadtalker_checkbox"):
+                with gr.TabItem('Settings'):
+                    with gr.Column(variant='panel'):
                         with gr.Row():
-                            source_image = gr.Image(label="Source image", source="upload", type="filepath", elem_id="img2img_image").style(width=512)
-
-                with gr.Tabs(elem_id="sadtalker_driven_audio"):
-                    with gr.TabItem('Upload OR TTS'):
-                        with gr.Column(variant='panel'):
-                            driven_audio = gr.Audio(label="Input audio", source="upload", type="filepath")
-
-                        if sys.platform != 'win32' and not in_webui: 
-                            from src.utils.text2speech import TTSTalker
-                            tts_talker = TTSTalker()
-                            with gr.Column(variant='panel'):
-                                input_text = gr.Textbox(label="Generating audio from text", lines=5, placeholder="please enter some text here, we genreate the audio from text using @Coqui.ai TTS.")
-                                tts = gr.Button('Generate audio',elem_id="sadtalker_audio_generate", variant='primary')
-                                tts.click(fn=tts_talker.test, inputs=[input_text], outputs=[driven_audio])
-                            
-            with gr.Column(variant='panel'): 
-                with gr.Tabs(elem_id="sadtalker_checkbox"):
-                    with gr.TabItem('Settings'):
-                        gr.Markdown("need help? please visit our [best practice page](https://github.com/OpenTalker/SadTalker/blob/main/docs/best_practice.md) for more detials")
-                        with gr.Column(variant='panel'):
-                            # width = gr.Slider(minimum=64, elem_id="img2img_width", maximum=2048, step=8, label="Manually Crop Width", value=512) # img2img_width
-                            # height = gr.Slider(minimum=64, elem_id="img2img_height", maximum=2048, step=8, label="Manually Crop Height", value=512) # img2img_width
-                            pose_style = gr.Slider(minimum=0, maximum=46, step=1, label="Pose style", value=0) # 
-                            size_of_image = gr.Radio([256, 512], value=256, label='face model resolution', info="use 256/512 model?") # 
-                            preprocess_type = gr.Radio(['crop', 'resize','full', 'extcrop', 'extfull'], value='crop', label='preprocess', info="How to handle input image?")
-                            is_still_mode = gr.Checkbox(label="Still Mode (fewer head motion, works with preprocess `full`)")
-                            batch_size = gr.Slider(label="batch size in generation", step=1, maximum=10, value=2)
-                            enhancer = gr.Checkbox(label="GFPGAN as Face enhancer")
-                            submit = gr.Button('Generate', elem_id="sadtalker_generate", variant='primary')
-                            
-                with gr.Tabs(elem_id="sadtalker_genearted"):
-                        gen_video = gr.Video(label="Generated video", format="mp4").style(width=256)
-
-        if warpfn:
-            submit.click(
-                        fn=warpfn(sad_talker.test), 
-                        inputs=[source_image,
-                                driven_audio,
-                                preprocess_type,
-                                is_still_mode,
-                                enhancer,
-                                batch_size,                            
-                                size_of_image,
-                                pose_style
-                                ], 
-                        outputs=[gen_video]
-                        )
-        else:
-            submit.click(
-                        fn=sad_talker.test, 
-                        inputs=[source_image,
-                                driven_audio,
-                                preprocess_type,
-                                is_still_mode,
-                                enhancer,
-                                batch_size,                            
-                                size_of_image,
-                                pose_style
-                                ], 
-                        outputs=[gen_video]
-                        )
-
-    return sadtalker_interface
- 
-
-if __name__ == "__main__":
-
-    demo = sadtalker_demo()
-    demo.queue()
-    demo.launch()
+                            pose_style = gr.Slider(
+                                minimum=0, maximum=45, step=1, label="Pose style", value=0)
+                            exp_weight = gr.Slider(
+                                minimum=0, maximum=3, step=0.1, label="expression scale", value=1)
+                            blink_every = gr.Checkbox(
+                                label="use eye blink", value=True)
+
+                        with gr.Row():
+                            size_of_image = gr.Radio(
+                                [256, 512], value=256, label='face model resolution', info="use 256/512 model?")
+                            preprocess_type = gr.Radio(
+                                ['crop', 'resize', 'full', 'extcrop', 'extfull'], value='crop', label='preprocess', info="How to handle input image?")
+
+                        with gr.Row():
+                            is_still_mode = gr.Checkbox(
+                                label="Still Mode (fewer head motion, works with preprocess `full`)")
+                            facerender = gr.Radio(
+                                ['facevid2vid', 'pirender'], value='facevid2vid', label='facerender', info="which face render?")
+
+                        with gr.Row():
+                            batch_size = gr.Slider(
+                                label="batch size in generation", step=1, maximum=10, value=1)
+                            enhancer = gr.Checkbox(
+                                label="GFPGAN as Face enhancer", value=True)
+
+                        submit = gr.Button(
+                            'Generate', elem_id="sadtalker_generate", variant='primary')
+
+            with gr.Tabs(elem_id="sadtalker_generated"):
+                gen_video = gr.Video(label="Generated video")
+
+    submit.click(
+        fn=generate_video,
+        inputs=[source_image, driven_audio, preprocess_type, is_still_mode, enhancer, batch_size, size_of_image,
+                pose_style, facerender, exp_weight, use_ref_video, ref_video, ref_info, use_idle_mode, length_of_audio, blink_every],
+        outputs=[gen_video],
+    )
 
+    with gr.Row():
+        gr.Examples(examples=get_examples(), inputs=[source_image, driven_audio, preprocess_type, is_still_mode, enhancer],
+                    outputs=[gen_video], fn=generate_video)
 
+demo.launch(debug=True)