load-base-model-first

VADER-VideoCrafter/scripts/main/train_t2v_lora.py

@@ -567,7 +567,7 @@ def should_sample(global_step, validation_steps, is_sample_preview):
     and is_sample_preview
 
 
-def run_training(args, peft_model, **kwargs):
+def run_training(args, model, **kwargs):
     ## ---------------------step 1: accelerator setup---------------------------
     accelerator = Accelerator(                                                  # Initialize Accelerator
         gradient_accumulation_steps=args.gradient_accumulation_steps,
@@ -576,6 +576,29 @@ def run_training(args, peft_model, **kwargs):
     )
     output_dir = args.project_dir
 
+
+    # step 2.1: add LoRA using peft
+    config = peft.LoraConfig(
+            r=args.lora_rank,
+            target_modules=["to_k", "to_v", "to_q"],        # only diffusion_model has these modules
+            lora_dropout=0.01,
+        )
+    
+    peft_model = peft.get_peft_model(model, config)
+
+    peft_model.print_trainable_parameters()
+
+    # load the pretrained LoRA model
+    if args.lora_ckpt_path != "Base Model":
+        if args.lora_ckpt_path == "huggingface-hps-aesthetic":  # download the pretrained LoRA model from huggingface
+            snapshot_download(repo_id='zheyangqin/VADER', local_dir ='VADER-VideoCrafter/checkpoints/pretrained_lora')
+            args.lora_ckpt_path = 'VADER-VideoCrafter/checkpoints/pretrained_lora/vader_videocrafter_hps_aesthetic.pt'
+        elif args.lora_ckpt_path == "huggingface-pickscore":    # download the pretrained LoRA model from huggingface
+            snapshot_download(repo_id='zheyangqin/VADER', local_dir ='VADER-VideoCrafter/checkpoints/pretrained_lora')
+            args.lora_ckpt_path = 'VADER-VideoCrafter/checkpoints/pretrained_lora/vader_videocrafter_pickscore.pt'
+        # load the pretrained LoRA model
+        peft.set_peft_model_state_dict(peft_model, torch.load(args.lora_ckpt_path))
+
     # Make one log on every process with the configuration for debugging.
     create_logging(logging, logger, accelerator)
     
@@ -698,7 +721,7 @@ def run_training(args, peft_model, **kwargs):
     # ==================================================================
 
             
-def setup_model(lora_ckpt_path="huggingface-pickscore", lora_rank=16):
+def setup_model():
     parser = get_parser()
     args = parser.parse_args()
 
@@ -721,41 +744,13 @@ def setup_model(lora_ckpt_path="huggingface-pickscore", lora_rank=16):
         model.first_stage_model = model.first_stage_model.half()
         model.cond_stage_model = model.cond_stage_model.half()
     
-    # step 2.1: add LoRA using peft
-    config = peft.LoraConfig(
-            r=lora_rank,
-            target_modules=["to_k", "to_v", "to_q"],        # only diffusion_model has these modules
-            lora_dropout=0.01,
-        )
     
-    peft_model = peft.get_peft_model(model, config)
-
-    peft_model.print_trainable_parameters()
-
-    # load the pretrained LoRA model
-    if lora_ckpt_path != "Base Model":
-        if lora_ckpt_path == "huggingface-hps-aesthetic":  # download the pretrained LoRA model from huggingface
-            snapshot_download(repo_id='zheyangqin/VADER', local_dir ='VADER-VideoCrafter/checkpoints/pretrained_lora')
-            lora_ckpt_path = 'VADER-VideoCrafter/checkpoints/pretrained_lora/vader_videocrafter_hps_aesthetic.pt'
-        elif lora_ckpt_path == "huggingface-pickscore":    # download the pretrained LoRA model from huggingface
-            snapshot_download(repo_id='zheyangqin/VADER', local_dir ='VADER-VideoCrafter/checkpoints/pretrained_lora')
-            lora_ckpt_path = 'VADER-VideoCrafter/checkpoints/pretrained_lora/vader_videocrafter_pickscore.pt'
-        elif lora_ckpt_path == "peft_model_532":
-            lora_ckpt_path = 'VADER-VideoCrafter/checkpoints/pretrained_lora/peft_model_532.pt'
-        elif lora_ckpt_path == "peft_model_548":
-            lora_ckpt_path = 'VADER-VideoCrafter/checkpoints/pretrained_lora/peft_model_548.pt'
-        elif lora_ckpt_path == "peft_model_536":
-            lora_ckpt_path = 'VADER-VideoCrafter/checkpoints/pretrained_lora/peft_model_536.pt'
-        elif lora_ckpt_path == "peft_model_400":
-            lora_ckpt_path = 'VADER-VideoCrafter/checkpoints/pretrained_lora/peft_model_400.pt'
-        # load the pretrained LoRA model
-        peft.set_peft_model_state_dict(peft_model, torch.load(lora_ckpt_path))
 
     print("Model setup complete!")
-    return peft_model
+    return model
 
 
-def main_fn(prompt, seed=200, height=320, width=512, unconditional_guidance_scale=12, ddim_steps=25, ddim_eta=1.0,
+def main_fn(prompt, lora_model, lora_rank, seed=200, height=320, width=512, unconditional_guidance_scale=12, ddim_steps=25, ddim_eta=1.0,
          frames=24, savefps=10, model=None):
 
     parser = get_parser()
@@ -765,6 +760,8 @@ def main_fn(prompt, seed=200, height=320, width=512, unconditional_guidance_scal
 
     # overwrite the default arguments
     args.prompt_str = prompt
+    args.lora_ckpt_path = lora_model
+    args.lora_rank = lora_rank
     args.seed = seed
     args.height = height
     args.width = width

app.py

@@ -2,6 +2,7 @@ import gradio as gr
 import os
 import spaces
 import sys
+from copy import deepcopy
 sys.path.append('./VADER-VideoCrafter/scripts/main')
 sys.path.append('./VADER-VideoCrafter/scripts')
 sys.path.append('./VADER-VideoCrafter')
@@ -19,24 +20,26 @@ examples = [
      "huggingface-pickscore", 16, 204, 384, 512, 12.0, 25, 1.0, 24, 10]
 ]
 
-model = None # Placeholder for model
+model = setup_model()
 
 @spaces.GPU(duration=70)
-def gradio_main_fn(prompt, seed, height, width, unconditional_guidance_scale, ddim_steps, ddim_eta,
+def gradio_main_fn(prompt, lora_model, lora_rank, seed, height, width, unconditional_guidance_scale, ddim_steps, ddim_eta,
                    frames, savefps):
     global model
     if model is None:
         return "Model is not loaded. Please load the model first."
     video_path = main_fn(prompt=prompt,
+                    lora_model=lora_model,
+                    lora_rank=int(lora_rank),
                     seed=int(seed),
                     height=int(height), 
-                    width=int(width), 
-                    unconditional_guidance_scale=float(unconditional_guidance_scale), 
-                    ddim_steps=int(ddim_steps), 
+                    width=int(width),
+                    unconditional_guidance_scale=float(unconditional_guidance_scale),
+                    ddim_steps=int(ddim_steps),
                     ddim_eta=float(ddim_eta), 
-                    frames=int(frames),  
+                    frames=int(frames),
                     savefps=int(savefps),
-                    model=model)
+                    model=deepcopy(model))
 
     return video_path
 
@@ -60,35 +63,6 @@ def update_dropdown(lora_rank):
     else: # 0
         return gr.update(value="Base Model")
 
-@spaces.GPU(duration=180)
-def setup_model_progress(lora_model, lora_rank):
-    global model
-
-    # Disable buttons and show loading indicator
-    yield (gr.update(interactive=False), gr.update(interactive=False), gr.update(interactive=False), "Loading model...")
-
-    model = setup_model(lora_model, lora_rank)  # Ensure you pass the necessary parameters to the setup_model function
-    
-    # Enable buttons after loading and update indicator
-    yield (gr.update(interactive=True), gr.update(interactive=True), gr.update(interactive=True), "Model loaded successfully")
-
-@spaces.GPU(duration=300)
-def generate_example(prompt, lora_model, lora_rank, seed, height, width, unconditional_guidance_scale, ddim_steps, ddim_eta,
-                   frames, savefps):
-    global model
-    model = setup_model(lora_model, lora_rank)
-    video_path = main_fn(prompt=prompt,
-                        seed=int(seed),
-                        height=int(height), 
-                        width=int(width), 
-                        unconditional_guidance_scale=float(unconditional_guidance_scale), 
-                        ddim_steps=int(ddim_steps), 
-                        ddim_eta=float(ddim_eta), 
-                        frames=int(frames),  
-                        savefps=int(savefps),
-                        model=model)
-    return video_path
-
 custom_css = """
     #centered {
         display: flex;
@@ -215,23 +189,19 @@ with gr.Blocks(css=custom_css) as demo:
                 value="huggingface-pickscore"
             )
             lora_rank = gr.Slider(minimum=8, maximum=16, label="LoRA Rank", step = 8, value=16)
-            load_btn = gr.Button("Load Model")
-            # Add a label to show the loading indicator
-            loading_indicator = gr.Label(value="", label="Loading Indicator")
+            prompt = gr.Textbox(placeholder="Enter prompt text here", lines=4, label="Text Prompt",
+                                value="A mermaid with flowing hair and a shimmering tail discovers a hidden underwater kingdom adorned with coral palaces, glowing pearls, and schools of colorful fish, encountering both wonders and dangers along the way.")
+            run_btn = gr.Button("Run Inference")
 
         with gr.Column(scale=0.3):
             output_video = gr.Video(elem_id="image-upload")
             
     with gr.Row(elem_id="centered"):
         with gr.Column(scale=0.6):      
-            prompt = gr.Textbox(placeholder="Enter prompt text here", lines=4, label="Text Prompt",
-                                value="A mermaid with flowing hair and a shimmering tail discovers a hidden underwater kingdom adorned with coral palaces, glowing pearls, and schools of colorful fish, encountering both wonders and dangers along the way.")
+           
 
             seed = gr.Slider(minimum=0, maximum=65536, label="Seed", step = 1, value=200)
 
-            run_btn = gr.Button("Run Inference")
-
-            
             with gr.Row():
                 height = gr.Slider(minimum=0, maximum=1024, label="Height", step = 16, value=384)
                 width = gr.Slider(minimum=0, maximum=1024, label="Width", step = 16, value=512)
@@ -252,10 +222,10 @@ with gr.Blocks(css=custom_css) as demo:
             reset_btn.click(fn=reset_fn, outputs=[prompt, seed, height, width, unconditional_guidance_scale, DDIM_Steps, DDIM_Eta, frames, lora_rank, savefps, lora_model])
                 
 
-
-            load_btn.click(fn=setup_model_progress, inputs=[lora_model, lora_rank], outputs=[load_btn, run_btn, reset_btn, loading_indicator])
             run_btn.click(fn=gradio_main_fn, 
-                        inputs=[prompt, seed, height, width, unconditional_guidance_scale, DDIM_Steps, DDIM_Eta, frames, savefps],
+                        inputs=[prompt, lora_model, lora_rank,
+                                seed, height, width, unconditional_guidance_scale, 
+                                DDIM_Steps, DDIM_Eta, frames, savefps],
                         outputs=output_video
                         )
             
@@ -263,9 +233,11 @@ with gr.Blocks(css=custom_css) as demo:
             lora_rank.change(fn=update_dropdown, inputs=lora_rank, outputs=lora_model)
 
             gr.Examples(examples=examples,
-                    inputs=[prompt, lora_model, lora_rank, seed, height, width, unconditional_guidance_scale, DDIM_Steps, DDIM_Eta, frames, savefps],
+                    inputs=[prompt, lora_model, lora_rank, seed, 
+                            height, width, unconditional_guidance_scale, 
+                            DDIM_Steps, DDIM_Eta, frames, savefps],
                     outputs=output_video,
-                    fn=generate_example,
+                    fn=gradio_main_fn,
                     run_on_click=False,
                     cache_examples="lazy",
                     )

app_bk.py

@@ -0,0 +1,273 @@
+import gradio as gr
+import os
+import spaces
+import sys
+sys.path.append('./VADER-VideoCrafter/scripts/main')
+sys.path.append('./VADER-VideoCrafter/scripts')
+sys.path.append('./VADER-VideoCrafter')
+
+
+from train_t2v_lora import main_fn, setup_model
+
+examples = [
+    ["A fairy tends to enchanted, glowing flowers.", 'huggingface-hps-aesthetic', 8, 400, 384, 512, 12.0, 25, 1.0, 24, 10],
+    ["A cat playing an electric guitar in a loft with industrial-style decor and soft, multicolored lights.", 'huggingface-hps-aesthetic', 8, 206, 384, 512, 12.0, 25, 1.0, 24, 10],
+    ["A raccoon playing a guitar under a blossoming cherry tree.", 'huggingface-hps-aesthetic', 8, 204, 384, 512, 12.0, 25, 1.0, 24, 10],
+    ["A mermaid with flowing hair and a shimmering tail discovers a hidden underwater kingdom adorned with coral palaces, glowing pearls, and schools of colorful fish, encountering both wonders and dangers along the way.", 
+     "huggingface-pickscore", 16, 205, 384, 512, 12.0, 25, 1.0, 24, 10],
+    ["A talking bird with shimmering feathers and a melodious voice leads an adventure to find a legendary treasure, guiding through enchanted forests, ancient ruins, and mystical challenges.",
+     "huggingface-pickscore", 16, 204, 384, 512, 12.0, 25, 1.0, 24, 10]
+]
+
+model = None # Placeholder for model
+
+@spaces.GPU(duration=70)
+def gradio_main_fn(prompt, seed, height, width, unconditional_guidance_scale, ddim_steps, ddim_eta,
+                   frames, savefps):
+    global model
+    if model is None:
+        return "Model is not loaded. Please load the model first."
+    video_path = main_fn(prompt=prompt,
+                    seed=int(seed),
+                    height=int(height), 
+                    width=int(width), 
+                    unconditional_guidance_scale=float(unconditional_guidance_scale), 
+                    ddim_steps=int(ddim_steps), 
+                    ddim_eta=float(ddim_eta), 
+                    frames=int(frames),  
+                    savefps=int(savefps),
+                    model=model)
+
+    return video_path
+
+def reset_fn():
+    return ("A mermaid with flowing hair and a shimmering tail discovers a hidden underwater kingdom adorned with coral palaces, glowing pearls, and schools of colorful fish, encountering both wonders and dangers along the way.", 
+            200, 384, 512, 12.0, 25, 1.0, 24, 16, 10, "huggingface-pickscore")
+
+def update_lora_rank(lora_model):
+    if lora_model == "huggingface-pickscore":
+        return gr.update(value=16)
+    elif lora_model == "huggingface-hps-aesthetic":
+        return gr.update(value=8)
+    else: # "Base Model"
+        return gr.update(value=8)
+
+def update_dropdown(lora_rank):
+    if lora_rank == 16:
+        return gr.update(value="huggingface-pickscore")
+    elif lora_rank == 8:
+        return gr.update(value="huggingface-hps-aesthetic")
+    else: # 0
+        return gr.update(value="Base Model")
+
+@spaces.GPU(duration=120)
+def setup_model_progress(lora_model, lora_rank):
+    global model
+
+    # Disable buttons and show loading indicator
+    yield (gr.update(interactive=False), gr.update(interactive=False), gr.update(interactive=False), "Loading model...")
+
+    model = setup_model(lora_model, lora_rank)  # Ensure you pass the necessary parameters to the setup_model function
+    
+    # Enable buttons after loading and update indicator
+    yield (gr.update(interactive=True), gr.update(interactive=True), gr.update(interactive=True), "Model loaded successfully")
+
+@spaces.GPU(duration=180)
+def generate_example(prompt, lora_model, lora_rank, seed, height, width, unconditional_guidance_scale, ddim_steps, ddim_eta,
+                   frames, savefps):
+    global model
+    model = setup_model(lora_model, lora_rank)
+    video_path = main_fn(prompt=prompt,
+                        seed=int(seed),
+                        height=int(height), 
+                        width=int(width), 
+                        unconditional_guidance_scale=float(unconditional_guidance_scale), 
+                        ddim_steps=int(ddim_steps), 
+                        ddim_eta=float(ddim_eta), 
+                        frames=int(frames),  
+                        savefps=int(savefps),
+                        model=model)
+    return video_path
+
+custom_css = """
+    #centered {
+        display: flex;
+        justify-content: center;
+    }
+    .column-centered {
+        display: flex;
+        flex-direction: column;
+        align-items: center;
+        width: 60%;
+    }
+    #image-upload {
+        flex-grow: 1;
+    }
+    #params .tabs {
+        display: flex;
+        flex-direction: column;
+        flex-grow: 1;
+    }
+    #params .tabitem[style="display: block;"] {
+        flex-grow: 1;
+        display: flex !important;
+    }
+    #params .gap {
+        flex-grow: 1;
+    }
+    #params .form {
+        flex-grow: 1 !important;
+    }
+    #params .form > :last-child{
+        flex-grow: 1;
+    }
+"""
+
+with gr.Blocks(css=custom_css) as demo:
+    with gr.Row():
+        with gr.Column():
+            gr.HTML(
+                """
+                <h1 style='text-align: center; font-size: 3.2em; margin-bottom: 0.5em; font-family: Arial, sans-serif; margin: 20px;'>
+                    Video Diffusion Alignment via Reward Gradient
+                </h1>
+                """
+            )
+            gr.HTML(
+                """
+                <style>
+                    body {
+                        font-family: Arial, sans-serif;
+                        text-align: center;
+                        margin: 50px;
+                    }
+                    a {
+                        text-decoration: none !important;
+                        color: black !important;
+                    }
+
+                </style>
+                <body>
+                <div style="font-size: 1.4em; margin-bottom: 0.5em; ">
+                    <a href="https://mihirp1998.github.io">Mihir Prabhudesai</a><sup>*</sup>&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;
+                    <a href="https://russellmendonca.github.io/">Russell Mendonca</a><sup>*</sup>&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;
+                    <a href="mailto: zheyangqin.qzy@gmail.com">Zheyang Qin</a><sup>*</sup>&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;
+                    <a href="https://www.cs.cmu.edu/~katef/">Katerina Fragkiadaki</a><sup></sup>&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;
+                    <a href="https://www.cs.cmu.edu/~dpathak/">Deepak Pathak</a><sup></sup>
+
+
+                </div>
+                <div style="font-size: 1.3em; font-style: italic;">
+                    Carnegie Mellon University
+                </div>
+                </body>
+                """
+            )
+            gr.HTML(
+                """
+                <head>
+                <link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.0.0-beta3/css/all.min.css">
+
+                <style>
+                .button-container {
+                    display: flex;
+                    justify-content: center;
+                    gap: 10px;
+                    margin-top: 10px;
+                }
+
+                .button-container a {
+                    display: inline-flex;
+                    align-items: center;
+                    padding: 10px 20px;
+                    border-radius: 30px;
+                    border: 1px solid #ccc;
+                    text-decoration: none;
+                    color: #333 !important;
+                    font-size: 16px;
+                    text-decoration: none !important;
+                }
+
+                .button-container a i {
+                    margin-right: 8px;
+                }
+                </style>
+                </head>
+
+                <div class="button-container">
+                <a href="https://arxiv.org/abs/2407.08737" class="btn btn-outline-primary">
+                    <i class="fa-solid fa-file-pdf"></i> Paper
+                </a>
+                <a href="https://vader-vid.github.io/" class="btn btn-outline-danger">
+                    <i class="fa-solid fa-video"></i> Website
+                <a href="https://github.com/mihirp1998/VADER" class="btn btn-outline-secondary">
+                    <i class="fa-brands fa-github"></i> Code
+                </a>
+                </div>
+                """
+            )
+
+    with gr.Row(elem_id="centered"):
+        with gr.Column(scale=0.3, elem_id="params"):
+            lora_model = gr.Dropdown(
+                label="VADER Model",
+                choices=["huggingface-pickscore", "huggingface-hps-aesthetic", "Base Model"],
+                value="huggingface-pickscore"
+            )
+            lora_rank = gr.Slider(minimum=8, maximum=16, label="LoRA Rank", step = 8, value=16)
+            load_btn = gr.Button("Load Model")
+            # Add a label to show the loading indicator
+            loading_indicator = gr.Label(value="", label="Loading Indicator")
+
+        with gr.Column(scale=0.3):
+            output_video = gr.Video(elem_id="image-upload")
+            
+    with gr.Row(elem_id="centered"):
+        with gr.Column(scale=0.6):      
+            prompt = gr.Textbox(placeholder="Enter prompt text here", lines=4, label="Text Prompt",
+                                value="A mermaid with flowing hair and a shimmering tail discovers a hidden underwater kingdom adorned with coral palaces, glowing pearls, and schools of colorful fish, encountering both wonders and dangers along the way.")
+
+            seed = gr.Slider(minimum=0, maximum=65536, label="Seed", step = 1, value=200)
+
+            run_btn = gr.Button("Run Inference")
+
+            
+            with gr.Row():
+                height = gr.Slider(minimum=0, maximum=1024, label="Height", step = 16, value=384)
+                width = gr.Slider(minimum=0, maximum=1024, label="Width", step = 16, value=512)
+
+            with gr.Row():
+                frames = gr.Slider(minimum=0, maximum=50, label="Frames", step = 1, value=24)
+                savefps = gr.Slider(minimum=0, maximum=60, label="Save FPS", step = 1, value=10)
+            
+            
+            with gr.Row():
+                DDIM_Steps = gr.Slider(minimum=0, maximum=100, label="DDIM Steps", step = 1, value=25)
+                unconditional_guidance_scale = gr.Slider(minimum=0, maximum=50, label="Guidance Scale", step = 0.1, value=12.0)
+                DDIM_Eta = gr.Slider(minimum=0, maximum=1, label="DDIM Eta", step = 0.01, value=1.0)
+
+            # reset button
+            reset_btn = gr.Button("Reset")
+            
+            reset_btn.click(fn=reset_fn, outputs=[prompt, seed, height, width, unconditional_guidance_scale, DDIM_Steps, DDIM_Eta, frames, lora_rank, savefps, lora_model])
+                
+
+
+            load_btn.click(fn=setup_model_progress, inputs=[lora_model, lora_rank], outputs=[load_btn, run_btn, reset_btn, loading_indicator])
+            run_btn.click(fn=gradio_main_fn, 
+                        inputs=[prompt, seed, height, width, unconditional_guidance_scale, DDIM_Steps, DDIM_Eta, frames, savefps],
+                        outputs=output_video
+                        )
+            
+            lora_model.change(fn=update_lora_rank, inputs=lora_model, outputs=lora_rank)
+            lora_rank.change(fn=update_dropdown, inputs=lora_rank, outputs=lora_model)
+
+            gr.Examples(examples=examples,
+                    inputs=[prompt, lora_model, lora_rank, seed, height, width, unconditional_guidance_scale, DDIM_Steps, DDIM_Eta, frames, savefps],
+                    outputs=output_video,
+                    fn=generate_example,
+                    run_on_click=False,
+                    cache_examples="lazy",
+                    )
+
+demo.launch(share=True)

gradio_cached_examples/32/indices.csv

@@ -0,0 +1 @@
+0

gradio_cached_examples/32/log.csv

@@ -0,0 +1,2 @@
+component 0,flag,username,timestamp
+"{""video"": {""path"": ""gradio_cached_examples/32/component 0/fd156c6a458fa048724e/temporal.mp4"", ""url"": ""/file=/tmp/gradio/4bc133becbc469de8da700250f7f7df1103c6f56/temporal.mp4"", ""size"": null, ""orig_name"": ""temporal.mp4"", ""mime_type"": null, ""is_stream"": false, ""meta"": {""_type"": ""gradio.FileData""}}, ""subtitles"": null}",,,2024-07-19 00:00:10.509808

gradio_cached_examples/34/indices.csv

@@ -0,0 +1 @@
+0

gradio_cached_examples/34/log.csv

@@ -0,0 +1,2 @@
+component 0,flag,username,timestamp
+"{""video"": {""path"": ""gradio_cached_examples/34/component 0/d2ac1c9664e80f60d50f/temporal.mp4"", ""url"": ""/file=/tmp/gradio/4bc133becbc469de8da700250f7f7df1103c6f56/temporal.mp4"", ""size"": null, ""orig_name"": ""temporal.mp4"", ""mime_type"": null, ""is_stream"": false, ""meta"": {""_type"": ""gradio.FileData""}}, ""subtitles"": null}",,,2024-07-18 23:33:26.912888