Create app.py

app.py

@@ -0,0 +1,166 @@
+import gradio as gr
+import torch
+from diffusers import MusicLDMPipeline
+
+
+# make Space compatible with CPU duplicates
+if torch.cuda.is_available():
+    device = "cuda"
+    torch_dtype = torch.float16
+else:
+    device = "cpu"
+    torch_dtype = torch.float32
+
+# load the diffusers pipeline
+pipe = MusicLDMPipeline.from_pretrained("cvssp/musicldm", torch_dtype=torch_dtype).to(device)
+
+# set the generator for reproducibility
+generator = torch.Generator(device)
+
+
+def text2audio(text, negative_prompt, duration, guidance_scale, random_seed, n_candidates):
+    if text is None:
+        raise gr.Error("Please provide a text input.")
+
+    waveforms = pipe(
+        text,
+        audio_length_in_s=duration,
+        guidance_scale=guidance_scale,
+        num_inference_steps=200,
+        negative_prompt=negative_prompt,
+        num_waveforms_per_prompt=n_candidates if n_candidates else 1,
+        generator=generator.manual_seed(int(random_seed)),
+    )["audios"]
+
+    return gr.make_waveform((16000, waveforms[0]), bg_image="bg.png")
+
+
+iface = gr.Blocks()
+
+with iface:
+    gr.HTML(
+        """
+            <div style="text-align: center; max-width: 700px; margin: 0 auto;">
+              <div
+                style="
+                  display: inline-flex; align-items: center; gap: 0.8rem; font-size: 1.75rem;
+                "
+              >
+                <h1 style="font-weight: 900; margin-bottom: 7px; line-height: normal;">
+                  MusicLDM: Enhancing Novelty in Text-to-Music Generation Using Beat-Synchronous Mixup Strategies
+                </h1>
+              </div> <p style="margin-bottom: 10px; font-size: 94%">
+                <a href="https://arxiv.org/abs/2308.01546">[Paper]</a> <a href="https://musicldm.github.io/">[Project
+                page]</a> <a href="https://huggingface.co/docs/diffusers/main/en/api/pipelines/musicldm">[🧨
+                Diffusers]</a>
+              </p>
+            </div>
+        """
+    )
+    gr.HTML("""This is the demo for MusicLDM, powered by 🧨 Diffusers. Demo uses the base checkpoint <a
+        href="https://huggingface.co/ircam-reach/musicldm"> ircam-reach/musicldm </a>. For faster inference without waiting in
+        queue, you may want to duplicate the space and upgrade to a GPU in the settings.""")
+    gr.DuplicateButton()
+
+    with gr.Group():
+        textbox = gr.Textbox(
+            value="Western music, chill out, folk instrument R & B beat",
+            max_lines=1,
+            label="Input text",
+            info="Your text is important for the audio quality. Please ensure it is descriptive by using more adjectives.",
+            elem_id="prompt-in",
+        )
+        negative_textbox = gr.Textbox(
+            value="low quality, average quality",
+            max_lines=1,
+            label="Negative prompt",
+            info="Enter a negative prompt not to guide the audio generation. Selecting appropriate negative prompts can improve the audio quality significantly.",
+            elem_id="prompt-in",
+        )
+
+        with gr.Accordion("Click to modify detailed configurations", open=False):
+            seed = gr.Number(
+                value=42,
+                label="Seed",
+                info="Change this value (any integer number) will lead to a different generation result.",
+            )
+            duration = gr.Slider(5, 15, value=10, step=2.5, label="Duration (seconds)")
+            guidance_scale = gr.Slider(
+                0,
+                7,
+                value=3.5,
+                step=0.5,
+                label="Guidance scale",
+                info="Larger => better quality and relevancy to text; Smaller => better diversity",
+            )
+            n_candidates = gr.Slider(
+                1,
+                5,
+                value=3,
+                step=1,
+                label="Number waveforms to generate",
+                info="Automatic quality control. This number control the number of candidates (e.g., generate three audios and choose the best to show you). A larger value usually lead to better quality with heavier computation",
+            )
+
+        outputs = gr.Video(label="Output", elem_id="output-video")
+        btn = gr.Button("Submit")
+
+    btn.click(
+        text2audio,
+        inputs=[textbox, negative_textbox, duration, guidance_scale, seed, n_candidates],
+        outputs=[outputs],
+    )
+
+    gr.HTML(
+        """
+    <div class="footer" style="text-align: center">
+        <p>Share your generations with the community by clicking the share icon at the top right the generated audio!</p>
+        <p>Follow the latest updates of MusicLDM on our<a href="https://musicldm.github.io/"
+        style="text-decoration: underline;" target="_blank"> project page </a> </p> 
+        <p>Model by <a
+        href="https://www.knutchen.com" style="text-decoration: underline;" target="_blank">Ke Chen</a>. Code and demo by 🤗 Hugging Face.</p>
+    </div>
+    """
+    )
+    gr.Examples(
+        [
+            ["Light rhythm techno", "low quality, average quality", 10, 3.5, 42, 3],
+            ["Futuristic drum and bass", "low quality, average quality", 10, 3.5, 42, 3],
+            ["Royal Film Music Orchestra", "low quality, average quality", 10, 3.5, 42, 3],
+            ["Elegant and gentle tunes of string quartet + harp", "low quality, average quality", 10, 3.5, 42, 3],
+            ["A fantastic piece of music with the deep sound of overlapping pianos", "low quality, average quality", 10, 3.5, 42, 3],
+            ["Gentle live acoustic guitar", "low quality, average quality", 10, 3.5, 42, 3],
+            ["Lyrical ballad played by saxophone", "low quality, average quality", 10, 3.5, 42, 3],
+        ],
+        fn=text2audio,
+        inputs=[textbox, negative_textbox, duration, guidance_scale, seed, n_candidates],
+        outputs=[outputs],
+        cache_examples=True,
+    )
+    gr.HTML(
+        """
+            <div class="acknowledgements"> <p>Essential Tricks for Enhancing the Quality of Your Generated
+            Audio</p>
+            <p>1. Try using more adjectives to describe your sound. For example: "Techno music with high melodic 
+            riffs and euphoric melody" is better than "Techno".</p>
+            <p>2. Try using different random seeds, which can significantly affect the quality of the generated 
+            output.</p>
+            <p>3. It's better to use general terms like 'techno' or 'jazz' instead of specific names for genres, 
+            artists or styles that the model may not be familiar with.</p>
+            <p>4. Using a negative prompt to not guide the diffusion process can improve the
+            audio quality significantly. Try using negative prompts like 'low quality'.</p>
+            </div>
+            """
+    )
+    with gr.Accordion("Additional information", open=False):
+        gr.HTML(
+            """
+            <div class="acknowledgments">
+                <p> We build the model with data from the <a href="https://audiostock.net//">Audiostock</a>,
+                dataset. The model is licensed as CC-BY-NC-4.0.
+                </p>
+            </div>
+            """
+        )
+
+iface.queue(max_size=20).launch()