Build Sulphur 2 Base demo

.gitignore

@@ -0,0 +1,4 @@
+__pycache__/
+*.py[cod]
+.DS_Store
+work/

README.md

@@ -1,13 +1,36 @@
 ---
-title: Sulphur 2 Base Demo
-emoji: 📚
-colorFrom: blue
-colorTo: purple
+title: Sulphur 2 Base
+emoji: 🎬
+colorFrom: purple
+colorTo: pink
 sdk: gradio
 sdk_version: 6.20.0
-python_version: '3.12'
 app_file: app.py
-pinned: false
+short_description: Generate short videos with Sulphur 2 Base
+python_version: "3.12"
+startup_duration_timeout: 1h
+models:
+  - SulphurAI/Sulphur-2-base
+  - diffusers/LTX-2.3-Diffusers
+tags:
+  - text-to-video
+  - ltx-2
+  - zero-gpu
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Sulphur 2 Base demo
+
+Generate short, synchronized audio-video clips with
+[SulphurAI/Sulphur-2-base](https://huggingface.co/SulphurAI/Sulphur-2-base),
+a fine-tune of Lightricks LTX 2.3.
+
+The demo uses the repository's FP8 mixed development checkpoint with the
+upstream LTX 2.3 Diffusers components. It runs a compact 512×320, 49-frame
+preset to keep interactive generation practical.
+
+## Responsible use
+
+Do not use this demo to create sexual content, content involving minors,
+non-consensual intimate imagery, graphic violence, or deceptive impersonation.
+Generated media may contain artifacts and should not be treated as factual.
+

app.py

@@ -0,0 +1,174 @@
+import os
+import re
+import tempfile
+import time
+
+os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
+
+import spaces
+import torch
+import gradio as gr
+from diffusers import AutoModel, LTX2Pipeline
+from diffusers.pipelines.ltx2.export_utils import encode_video
+from diffusers.pipelines.ltx2.utils import DEFAULT_NEGATIVE_PROMPT
+
+
+MODEL_ID = "SulphurAI/Sulphur-2-base"
+BASE_MODEL_ID = "diffusers/LTX-2.3-Diffusers"
+CHECKPOINT_URL = (
+    "https://huggingface.co/SulphurAI/Sulphur-2-base/"
+    "blob/main/sulphur_dev_fp8mixed.safetensors"
+)
+
+WIDTH = 512
+HEIGHT = 320
+NUM_FRAMES = 49
+FPS = 24.0
+
+# A deliberately conservative public-demo guard. The upstream checkpoint is
+# described as uncensored, but a public demo should not generate abuse content.
+BLOCKED_PATTERNS = (
+    r"\b(child|children|kid|minor|underage|teen(?:ager)?)\b.{0,50}"
+    r"\b(nude|naked|sex|sexual|explicit|porn)\b",
+    r"\b(nude|naked|sex|sexual|explicit|porn)\b.{0,50}"
+    r"\b(child|children|kid|minor|underage|teen(?:ager)?)\b",
+    r"\b(rape|sexual assault|non[- ]consensual|revenge porn|csam)\b",
+    r"\b(gore|dismemberment|beheading|graphic violence)\b",
+)
+
+
+def _allowed(prompt: str) -> bool:
+    text = prompt.casefold()
+    return not any(re.search(pattern, text) for pattern in BLOCKED_PATTERNS)
+
+
+transformer = AutoModel.from_single_file(
+    CHECKPOINT_URL,
+    torch_dtype=torch.bfloat16,
+)
+pipe = LTX2Pipeline.from_pretrained(
+    BASE_MODEL_ID,
+    transformer=transformer,
+    torch_dtype=torch.bfloat16,
+).to("cuda")
+pipe.vae.enable_tiling()
+
+
+def _duration(prompt: str, seed: int, steps: int, *args, **kwargs) -> int:
+    del prompt, seed, args, kwargs
+    return min(300, 90 + int(steps) * 9)
+
+
+@spaces.GPU(duration=_duration, size="xlarge")
+def generate(prompt: str, seed: int, steps: int) -> tuple[str, str]:
+    """Generate a short 512×320 video with synchronized audio from a text prompt."""
+    prompt = (prompt or "").strip()
+    if len(prompt) < 8:
+        raise gr.Error("Please enter a more descriptive prompt.")
+    if len(prompt) > 1_500:
+        raise gr.Error("Please keep the prompt under 1,500 characters.")
+    if not _allowed(prompt):
+        raise gr.Error(
+            "This public demo cannot process sexual, exploitative, or graphic-violence prompts."
+        )
+
+    started = time.perf_counter()
+    generator = torch.Generator(device="cuda").manual_seed(int(seed))
+    video, audio = pipe(
+        prompt=prompt,
+        negative_prompt=DEFAULT_NEGATIVE_PROMPT,
+        width=WIDTH,
+        height=HEIGHT,
+        num_frames=NUM_FRAMES,
+        frame_rate=FPS,
+        num_inference_steps=int(steps),
+        guidance_scale=3.0,
+        generator=generator,
+        output_type="np",
+        return_dict=False,
+    )
+
+    output = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False)
+    output.close()
+    encode_video(
+        video[0],
+        fps=FPS,
+        audio=audio[0].float().cpu(),
+        audio_sample_rate=pipe.vocoder.config.output_sampling_rate,
+        output_path=output.name,
+    )
+    elapsed = time.perf_counter() - started
+    return output.name, f"Finished in {elapsed:.1f}s · seed {int(seed)}"
+
+
+CSS = """
+.gradio-container { max-width: 1120px !important; }
+.hero { text-align: center; margin: 1.5rem auto 1rem; }
+.hero h1 { font-size: clamp(2rem, 5vw, 4rem); margin-bottom: .2rem; }
+.hero p { color: #a1a1aa; font-size: 1.05rem; }
+"""
+
+with gr.Blocks(css=CSS, theme=gr.themes.Soft(primary_hue="purple")) as demo:
+    gr.HTML(
+        """
+        <div class="hero">
+          <h1>🎬 Sulphur 2 Base</h1>
+          <p>Text-to-video with synchronized audio, powered by LTX 2.3.</p>
+        </div>
+        """
+    )
+    with gr.Row():
+        with gr.Column(scale=5):
+            prompt = gr.Textbox(
+                label="Describe your shot",
+                placeholder=(
+                    "A cinematic tracking shot of a tiny moss-covered robot "
+                    "walking through a rain-soaked neon market..."
+                ),
+                lines=7,
+                max_lines=12,
+            )
+            with gr.Row():
+                seed = gr.Number(label="Seed", value=42, precision=0)
+                steps = gr.Slider(
+                    label="Inference steps", minimum=12, maximum=30, value=20, step=1
+                )
+            run = gr.Button("Generate video", variant="primary", size="lg")
+            gr.Markdown(
+                "Public demo guardrails apply. Avoid sexual, exploitative, "
+                "graphic, or deceptive content."
+            )
+        with gr.Column(scale=7):
+            video = gr.Video(label="Generated clip", autoplay=True)
+            status = gr.Markdown()
+
+    gr.Examples(
+        examples=[
+            [
+                "A macro cinematic shot of a glass terrarium at dawn. A tiny "
+                "clockwork hummingbird unfolds its brass wings, dew glints on "
+                "fern leaves, soft mechanical clicks and distant birdsong."
+            ],
+            [
+                "A wide aerial shot over black volcanic sand at blue hour. "
+                "Bioluminescent waves roll ashore under a star-filled sky, with "
+                "wind and gentle surf in the soundtrack."
+            ],
+            [
+                "Stop-motion style: a paper astronaut plants a small sunflower "
+                "on a handcrafted moon, warm studio lighting, subtle paper "
+                "rustling and a whimsical music-box melody."
+            ],
+        ],
+        inputs=[prompt],
+        cache_examples=False,
+    )
+    run.click(
+        fn=generate,
+        inputs=[prompt, seed, steps],
+        outputs=[video, status],
+        api_name="generate",
+    )
+
+demo.queue(default_concurrency_limit=1).launch(mcp_server=True)
+

requirements.txt

@@ -0,0 +1,7 @@
+git+https://github.com/huggingface/diffusers.git
+transformers>=4.57.0
+accelerate>=1.10.0
+safetensors>=0.6.2
+sentencepiece>=0.2.1
+imageio-ffmpeg>=0.6.0
+