Upload folder using huggingface_hub

README.md

@@ -12,9 +12,10 @@ license: mit
 short_description: Creative ACE-Step 1.5 text-to-song with genre-first UI
 suggested_hardware: zero-a10g
 models:
-  - ACE-Step/acestep-v15-sft
-  - ACE-Step/acestep-v15-base
+  - Runware/acestep-v15-turbo-diffusers
   - ACE-Step/acestep-v15-xl-turbo-diffusers
+  - ACE-Step/acestep-v15-xl-sft-diffusers
+  - ACE-Step/acestep-v15-xl-base-diffusers
 tags:
   - music
   - text-to-audio
@@ -31,16 +32,19 @@ Official demos expose many technical controls and can feel unreliable. This Spac
 
 - Guides you with **genre → style → mood → tempo → lyrics**
 - Keeps expert diffusion settings collapsed
-- Defaults to the **non-XL SFT** checkpoint, with **XL Turbo** available for faster ZeroGPU runs
+- Defaults to **non-XL Turbo** in Diffusers format (ZeroGPU-friendly)
 - Treats **lyric generation as the primary path** (editable afterward)
 
 ## Models
 
+Only Diffusers `AceStepPipeline` checkpoints are supported:
+
 | UI label | Repo |
 |----------|------|
-| SFT (default) | `ACE-Step/acestep-v15-sft` |
-| Base | `ACE-Step/acestep-v15-base` |
+| Turbo non-XL (default) | `Runware/acestep-v15-turbo-diffusers` |
 | XL Turbo | `ACE-Step/acestep-v15-xl-turbo-diffusers` |
+| XL SFT | `ACE-Step/acestep-v15-xl-sft-diffusers` |
+| XL Base | `ACE-Step/acestep-v15-xl-base-diffusers` |
 
 Built with [`diffusers.AceStepPipeline`](https://huggingface.co/docs/diffusers/api/pipelines/ace_step).
 

app.py

@@ -52,7 +52,13 @@ def _load_pipe(repo_id: str) -> AceStepPipeline:
         return _pipes[repo_id]
     # Keep only one heavy pipeline in memory on ZeroGPU
     _pipes.clear()
-    pipe = AceStepPipeline.from_pretrained(repo_id, torch_dtype=torch.bfloat16)
+    try:
+        pipe = AceStepPipeline.from_pretrained(repo_id, torch_dtype=torch.bfloat16)
+    except Exception as e:
+        raise gr.Error(
+            f"Failed to load model `{repo_id}`.\n"
+            f"This Space only supports Diffusers AceStepPipeline checkpoints.\n\n{e}"
+        ) from e
     if hasattr(pipe, "vae") and hasattr(pipe.vae, "enable_tiling"):
         pipe.vae.enable_tiling()
     _pipes[repo_id] = pipe
@@ -64,6 +70,16 @@ def _model_cfg(model_name: str) -> dict:
     return MODELS.get(model_name, MODELS[DEFAULT_MODEL])
 
 
+# Preload default checkpoint on CPU during startup so ZeroGPU time isn't spent downloading.
+try:
+    _DEFAULT_REPO = MODELS[DEFAULT_MODEL]["repo_id"]
+    print(f"[startup] Preloading {_DEFAULT_REPO} on CPU…")
+    _load_pipe(_DEFAULT_REPO)
+    print("[startup] Default model ready")
+except Exception as e:
+    print(f"[startup] Default model preload skipped: {e}")
+
+
 def apply_genre(genre: str):
     """Fill creative defaults from genre (all still user-overridable)."""
     g = GENRES[genre]
@@ -221,8 +237,12 @@ def _generate_impl(
         task_type="text2music",
     )
 
-    output = pipe(**kwargs)
-    audio = output.audios[0]
+    try:
+        output = pipe(**kwargs)
+        audio = output.audios[0]
+    except Exception as e:
+        raise gr.Error(f"Generation failed:\n{type(e).__name__}: {e}") from e
+
     if isinstance(audio, torch.Tensor):
         audio = audio.detach().float().cpu().numpy()
     if audio.ndim == 2:
@@ -273,7 +293,7 @@ with gr.Blocks(title="ACE-Step Inspire", theme=gr.themes.Soft(primary_hue="green
             choices=list(MODELS.keys()),
             value=DEFAULT_MODEL,
             label="Model",
-            info="Default is non-XL SFT. Switch to XL Turbo if ZeroGPU queues feel slow.",
+            info="Default is non-XL Turbo (Diffusers). XL SFT is slower but often richer.",
         )
         duration = gr.Number(
             value=DEFAULT_DURATION,
@@ -416,8 +436,8 @@ with gr.Blocks(title="ACE-Step Inspire", theme=gr.themes.Soft(primary_hue="green
         """
 ---
 **Tips:** Start with **Generate lyrics** → **Build prompt** → **Generate song**.  
-For long tracks on ZeroGPU, prefer **XL Turbo**. Non-XL SFT is the quality default.  
-Lyrics here use a creative composer tuned to genre/mood (fast & reliable). ACE-Step’s own 5Hz planner LM is part of the heavy full stack and is intentionally not loaded in this Space.
+Default model is **non-XL Turbo** (works reliably on ZeroGPU). Use **XL SFT** for quality when you can wait.  
+Lyrics use a creative composer tuned to genre/mood. ACE-Step’s 5Hz planner LM is not loaded here (keeps the Space lean).
 """
     )
 

presets.py

@@ -475,33 +475,43 @@ VOCAL_LANGUAGES: dict[str, str] = {
     "Russian": "ru",
 }
 
+# Only Diffusers-format AceStepPipeline checkpoints work in this Space.
+# Native transformers checkpoints (acestep-v15-sft / base / Ace-Step1.5) need the full ACE-Step package.
 MODELS: dict[str, dict] = {
-    "SFT (quality, non-XL) — default": {
-        "repo_id": "ACE-Step/acestep-v15-sft",
+    "Turbo (non-XL) — default": {
+        "repo_id": "Runware/acestep-v15-turbo-diffusers",
+        "steps": 8,
+        "guidance": 1.0,
+        "shift": 3.0,
+        "turbo": True,
+        "gpu_seconds": 180,
+    },
+    "XL Turbo (faster / larger)": {
+        "repo_id": "ACE-Step/acestep-v15-xl-turbo-diffusers",
+        "steps": 8,
+        "guidance": 1.0,
+        "shift": 3.0,
+        "turbo": True,
+        "gpu_seconds": 200,
+    },
+    "XL SFT (higher quality, slower)": {
+        "repo_id": "ACE-Step/acestep-v15-xl-sft-diffusers",
         "steps": 28,
         "guidance": 7.0,
         "shift": 3.0,
         "turbo": False,
-        "gpu_seconds": 300,
+        "gpu_seconds": 360,
     },
-    "Base (more diversity, non-XL)": {
-        "repo_id": "ACE-Step/acestep-v15-base",
+    "XL Base (more diversity, slower)": {
+        "repo_id": "ACE-Step/acestep-v15-xl-base-diffusers",
         "steps": 32,
         "guidance": 7.0,
         "shift": 3.0,
         "turbo": False,
-        "gpu_seconds": 320,
-    },
-    "XL Turbo (fast, ZeroGPU-friendly)": {
-        "repo_id": "ACE-Step/acestep-v15-xl-turbo-diffusers",
-        "steps": 8,
-        "guidance": 1.0,
-        "shift": 3.0,
-        "turbo": True,
-        "gpu_seconds": 180,
+        "gpu_seconds": 360,
     },
 }
 
-DEFAULT_MODEL = "SFT (quality, non-XL) — default"
+DEFAULT_MODEL = "Turbo (non-XL) — default"
 DEFAULT_GENRE = "Pop"
 DEFAULT_DURATION = 180

requirements.txt

@@ -2,7 +2,8 @@
 torch
 torchaudio
 torchvision
-diffusers>=0.37.0
+# AceStepPipeline needs a recent Diffusers (0.38+)
+diffusers>=0.38.0
 transformers>=4.51.0
 accelerate>=1.12.0
 safetensors