Support custom inference lengths

README.md

@@ -21,8 +21,8 @@ tags:
   <a href="#code">
     <img src="https://img.shields.io/badge/Conference-ECCV%202026-orange" alt="Conference">
   </a>
-  <a href="#inference">
-    <img src="https://img.shields.io/badge/Inference-Custom_Music-8a2be2" alt="Inference">
+  <a href="https://huggingface.co/xlt99/FlowerDance">
+    <img src="https://img.shields.io/badge/Hugging_Face-%F0%9F%A4%97_Model-FFD21E" alt="Hugging Face">
   </a>
 </p>
 
@@ -34,8 +34,10 @@ tags:
 >
 > FlowerDance combines MeanFlow with Physical Consistency Constraints for high-quality few-step sampling, and uses a lightweight non-autoregressive BiMamba backbone with Channel-Level Fusion for long-horizon music-to-dance synthesis. It also supports motion editing through time-decayed soft masking, enabling users to refine generated dance sequences interactively.
 
-🎉 **FlowerDance has been accepted to ECCV 2026!**
-✨ Training and inference code released! ✨
+<p align="center">
+  <strong>🎉 FlowerDance has been accepted to ECCV 2026! 🎉</strong><br>
+  <em>✨ Training and inference code are now available. ✨</em>
+</p>
 
 ---
 
@@ -107,60 +109,37 @@ accelerate launch train.py --batch_size 128  --epochs 4000 --feature_type baseli
 ```
 ---
 
-<a id="inference"></a>
+## 📏 Evaluation
 
-## 🎵 Inference
+### 🧪 Evaluate the Model
 
-Generate a 3D dance sequence directly from your own music:
+To evaluate the our model’s performance:
 
 ```bash
-python inference.py path/to/music.wav \
-    --genre Hiphop \
-    --duration 32 \
-    --output-dir inference_outputs
+python test.py --batch_size 128
 ```
 
-The input can be a WAV, MP3, FLAC, or OGG file. The script extracts the same 35-dimensional baseline music features used during training, supports genre-conditioned generation, and automatically downloads `train-3700.pt` from [Hugging Face](https://huggingface.co/xlt99/FlowerDance) when `--checkpoint` is omitted.
+---
 
-List all 16 supported dance genres:
+<a id="inference"></a>
 
-```bash
-python inference.py --list-genres
-```
+## 🎵 Inference
 
-Generate a 60-second sequence from a selected part of a longer track:
+Generate a genre-conditioned dance sequence of a custom length from your own music:
 
 ```bash
-python inference.py path/to/music.mp3 \
-    --genre Popping \
-    --start 10 \
-    --duration 60
-```
-
-The default duration is 32 seconds and the maximum supported duration is 60 seconds. A CUDA GPU is required. Each run saves:
-
-```text
-inference_outputs/
-├── music_Hiphop.wav
-├── music_Hiphop_features.npy
-└── 0/
-    └── music_Hiphop.pkl
+python inference.py path/to/music.wav \
+    --genre Hiphop \
+    --duration 32
 ```
 
-The PKL file contains `smpl_poses`, `smpl_trans`, and `full_pose` for downstream visualization or rendering.
-
----
-
-## 📏 Evaluation
-
-### 🧪 Evaluate the Model
-
-To evaluate the our model’s performance:
+List all supported dance genres:
 
 ```bash
-python test.py --batch_size 128
+python inference.py --list-genres
 ```
 
+The checkpoint is downloaded automatically from [Hugging Face](https://huggingface.co/xlt99/FlowerDance) when `--checkpoint` is omitted. Generated motions are saved to `inference_outputs/`.
 
 ---
 

inference.py

@@ -18,7 +18,6 @@ SAMPLE_RATE = FPS * HOP_LENGTH
 FEATURE_DIM = 35
 MOTION_DIM = 151
 DEFAULT_DURATION = 32.0
-MAX_DURATION = 60.0
 
 GENRES = (
     "Dai",
@@ -426,8 +425,8 @@ def main():
     music_path = args.music.expanduser().resolve()
     if not music_path.is_file():
         parser.error(f"Music file not found: {music_path}")
-    if not 0 < args.duration <= MAX_DURATION:
-        parser.error(f"--duration must be in (0, {MAX_DURATION:g}] seconds")
+    if args.duration <= 0:
+        parser.error("--duration must be positive")
     if args.steps < 2:
         parser.error("--steps must be at least 2")