Spaces:
Paused
Paused
Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,107 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import torch
|
| 3 |
+
import gradio as gr
|
| 4 |
+
import spaces # Critical for ZeroGPU support in Spaces
|
| 5 |
+
from diffusers import DiffusionPipeline # Common for HF video models
|
| 6 |
+
from PIL import Image
|
| 7 |
+
|
| 8 |
+
# =============================================================
|
| 9 |
+
# INITIALIZATION & PIPELINE CONFIGURATION
|
| 10 |
+
# =============================================================
|
| 11 |
+
MODEL_ID = "MiniMaxAI/MiniMax-M1"
|
| 12 |
+
|
| 13 |
+
def load_pipeline():
|
| 14 |
+
# Load model with bfloat16 for modern GPU efficiency [17]
|
| 15 |
+
# Note: Ensure the model is available on the HF Hub or adjust path
|
| 16 |
+
pipe = DiffusionPipeline.from_pretrained(
|
| 17 |
+
MODEL_ID,
|
| 18 |
+
torch_dtype=torch.bfloat16,
|
| 19 |
+
use_safetensors=True
|
| 20 |
+
)
|
| 21 |
+
|
| 22 |
+
# Critical VRAM optimizations for deployment on 24GB or ZeroGPU [2]
|
| 23 |
+
pipe.enable_model_cpu_offload()
|
| 24 |
+
pipe.enable_vae_slicing()
|
| 25 |
+
return pipe
|
| 26 |
+
|
| 27 |
+
# Pipeline is initialized globally to avoid reloads on every click
|
| 28 |
+
pipe = load_pipeline()
|
| 29 |
+
|
| 30 |
+
# =============================================================
|
| 31 |
+
# GENERATION LOGIC
|
| 32 |
+
# =============================================================
|
| 33 |
+
@spaces.GPU(duration=300) # Allocated GPU time for complex generation [16]
|
| 34 |
+
def generate_video(prompt, negative_prompt, steps, guidance_scale, seed):
|
| 35 |
+
# Standard 5-second video length at 16 FPS [18-20]
|
| 36 |
+
num_frames = 81
|
| 37 |
+
|
| 38 |
+
# Use Generator for deterministic results [21, 22]
|
| 39 |
+
generator = torch.Generator("cuda").manual_seed(int(seed))
|
| 40 |
+
|
| 41 |
+
# Execution using VACE-style inputs [8, 23]
|
| 42 |
+
output = pipe(
|
| 43 |
+
prompt=prompt,
|
| 44 |
+
negative_prompt=negative_prompt,
|
| 45 |
+
num_inference_steps=int(steps),
|
| 46 |
+
guidance_scale=float(guidance_scale),
|
| 47 |
+
num_frames=num_frames,
|
| 48 |
+
generator=generator
|
| 49 |
+
).frames
|
| 50 |
+
|
| 51 |
+
# Exporting generated frames to a video file [24, 25]
|
| 52 |
+
# For a real app, use diffusers.utils.export_to_video
|
| 53 |
+
from diffusers.utils import export_to_video
|
| 54 |
+
import tempfile
|
| 55 |
+
|
| 56 |
+
temp_path = tempfile.mktemp(suffix=".mp4")
|
| 57 |
+
export_to_video(output, temp_path, fps=16)
|
| 58 |
+
return temp_path
|
| 59 |
+
|
| 60 |
+
# =============================================================
|
| 61 |
+
# UI DESIGN (Using gr.Blocks for Professional Layout)
|
| 62 |
+
# =============================================================
|
| 63 |
+
# CSS for custom styling to improve "Joyful Experience" [26, 27]
|
| 64 |
+
css = """
|
| 65 |
+
.container { max-width: 1000px; margin: auto; }
|
| 66 |
+
.gen_btn { background-color: #7224f2 !important; color: white !important; }
|
| 67 |
+
"""
|
| 68 |
+
|
| 69 |
+
with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
|
| 70 |
+
gr.Markdown(f"# {MODEL_ID} High-Fidelity Video Generator")
|
| 71 |
+
gr.Markdown("Leveraging ZeroGPU and VRAM offloading for cinematic AI video [2, 28].")
|
| 72 |
+
|
| 73 |
+
with gr.Row():
|
| 74 |
+
with gr.Column(scale=1):
|
| 75 |
+
prompt_input = gr.Textbox(
|
| 76 |
+
label="Prompt",
|
| 77 |
+
placeholder="Describe the action and scene details...",
|
| 78 |
+
lines=4
|
| 79 |
+
)
|
| 80 |
+
neg_prompt = gr.Textbox(
|
| 81 |
+
label="Negative Prompt",
|
| 82 |
+
value="blurry, distorted, low quality, watermark, text"
|
| 83 |
+
)
|
| 84 |
+
|
| 85 |
+
with gr.Accordion("Advanced Settings", open=False):
|
| 86 |
+
steps = gr.Slider(20, 50, value=30, step=1, label="Inference Steps")
|
| 87 |
+
guidance = gr.Slider(1.0, 15.0, value=7.0, label="Guidance Scale")
|
| 88 |
+
seed = gr.Number(value=42, label="Seed")
|
| 89 |
+
|
| 90 |
+
generate_btn = gr.Button("Generate Video", variant="primary", elem_classes="gen_btn")
|
| 91 |
+
|
| 92 |
+
with gr.Column(scale=1):
|
| 93 |
+
video_output = gr.Video(label="Generated Output")
|
| 94 |
+
|
| 95 |
+
# Event listener mapping [29, 30]
|
| 96 |
+
generate_btn.click(
|
| 97 |
+
fn=generate_video,
|
| 98 |
+
inputs=[prompt_input, neg_prompt, steps, guidance, seed],
|
| 99 |
+
outputs=video_output
|
| 100 |
+
)
|
| 101 |
+
|
| 102 |
+
# =============================================================
|
| 103 |
+
# LAUNCH
|
| 104 |
+
# =============================================================
|
| 105 |
+
if __name__ == "__main__":
|
| 106 |
+
# Ensure app.py is at the root for automatic HF detection [14, 15]
|
| 107 |
+
demo.launch()
|