Fix the input issue of the reference image

One-to-All-Animation/video-generation/inference_1.3b.py

@@ -76,6 +76,26 @@ def resolve_pose_video_path(pose_identifier: str) -> str:
     )
 
 
+def resolve_ref_image_path(ref_identifier: str) -> str:
+    """Return an absolute ref image path, searching ref_img_dict when needed."""
+    if not ref_identifier:
+        raise ValueError("ref_identifier is required to resolve a reference image path")
+
+    candidates = []
+    expanded = os.path.expanduser(ref_identifier)
+    candidates.append(expanded)
+    if not os.path.isabs(expanded):
+        candidates.append(os.path.join(ref_image_dir, ref_identifier))
+
+    for candidate in candidates:
+        if candidate and os.path.exists(candidate):
+            return os.path.abspath(candidate)
+
+    raise FileNotFoundError(
+        f"❌ Reference image not found. Tried: {', '.join(os.path.abspath(c) for c in candidates)}"
+    )
+
+
 def resizecrop(image, th, tw):
     w, h = image.size
     if h / w > th / tw:
@@ -371,7 +391,8 @@ def unified_worker(rank, world_size, task_list, ckpt_path, use_direct_paths=Fals
     torch.cuda.empty_cache()
 
 
-def run_all_tasks(ckpt_path, img_g, pose_g, cache_dir_root="input_cache2", use_direct_paths=False, pose_video_override=None):
+def run_all_tasks(ckpt_path, img_g, pose_g, cache_dir_root="input_cache2",
+                  use_direct_paths=False, pose_video_override=None, ref_image_override=None):
     """扫描 cache 目录，跑全部任务"""
     # Modified: Direct paths mode for StableSigner3 generated videos
     # Format: (ref_image_path, pose_video_path, new_h, new_w, ref_cfg, pose_cfg, prompt)
@@ -391,9 +412,17 @@ def run_all_tasks(ckpt_path, img_g, pose_g, cache_dir_root="input_cache2", use_d
             if not pose_cfg: pose_cfg = pose_g
             else:            pose_cfg = float(pose_cfg)
 
-            if not os.path.exists(ref_path):
-                print(f"❌ Reference image not found: {ref_path}")
-                continue
+            if ref_image_override:
+                try:
+                    ref_target = resolve_ref_image_path(ref_image_override)
+                except (FileNotFoundError, ValueError) as exc:
+                    print(exc)
+                    continue
+            else:
+                ref_target = ref_path
+                if not os.path.exists(ref_target):
+                    print(f"❌ Reference image not found: {ref_target}")
+                    continue
 
             pose_target = pose_video_override if pose_video_override else pose_video_path
             try:
@@ -402,13 +431,13 @@ def run_all_tasks(ckpt_path, img_g, pose_g, cache_dir_root="input_cache2", use_d
                 print(exc)
                 continue
 
-            refname = os.path.splitext(os.path.basename(ref_path))[0]
+            refname = os.path.splitext(os.path.basename(ref_target))[0]
             vidname = os.path.splitext(os.path.basename(resolved_pose_path))[0]
             job_name = f"ref_{refname}_vid_{vidname}"
 
             # Create a simple dict to hold paths
             task_info = {
-                'ref_path': ref_path,
+                'ref_path': ref_target,
                 'pose_video_path': resolved_pose_path,
                 'job_name': job_name,
                 'new_h': new_h,
@@ -469,6 +498,12 @@ if __name__ == "__main__":
         default=None,
         help="Pose video path or filename inside generated_pose_video"
     )
+    parser.add_argument(
+        "--ref_image",
+        type=str,
+        default=None,
+        help="Reference image path (absolute or relative). Defaults to square2.jpg"
+    )
     parser.add_argument(
         "--use_cache_mode",
         action="store_true",
@@ -488,7 +523,8 @@ if __name__ == "__main__":
                     img_g,
                     pose_g,
                     use_direct_paths=True,
-                    pose_video_override=args.pose_video
+                    pose_video_override=args.pose_video,
+                    ref_image_override=args.ref_image
                 )
             else:
                 print("📁 使用缓存模式: 从 input_cache 读取")

One-to-All-Animation/video-generation/run_batch.sh

@@ -0,0 +1,17 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+source /research/cbim/vast/sf895/miniforge3/bin/activate one-to-all
+
+cd "$(dirname "$0")"
+
+#cd /research/cbim/vast/sf895/code/StableSigner/One-to-All-Animation/video-generation
+#conda activate one-to-all
+CUDA_VISIBLE_DEVICES=0 python inference_1.3b.py --pose_video HELLO_GOOD_MORNING_20251215_142658_openpose_style.mp4 --ref_image square1.png
+CUDA_VISIBLE_DEVICES=1 python inference_1.3b.py --pose_video HELLO_WORLD_20251215_142017_openpose_style.mp4 --ref_image square2.png
+CUDA_VISIBLE_DEVICES=2 python inference_1.3b.py --pose_video I_GO_SCHOOL_LATER_20251215_142204_openpose_style.mp4 --ref_image square3.png
+CUDA_VISIBLE_DEVICES=3 python inference_1.3b.py --pose_video I_NEED_HELP_PLEASE_20251215_142045_openpose_style.mp4 --ref_image square4.png
+CUDA_VISIBLE_DEVICES=4 python inference_1.3b.py --pose_video TODAY_WEATHER_NICE_20251215_142126_openpose_style.mp4 --ref_image square5.png
+CUDA_VISIBLE_DEVICES=5 python inference_1.3b.py --pose_video WEHRE_IS_MY_CAT_20251215_142307_openpose_style.mp4 --ref_image square6.png
+CUDA_VISIBLE_DEVICES=6 python inference_1.3b.py --pose_video TODAY_WEATHER_NICE_20251215_142126_openpose_style.mp4 --ref_image square2.png
+CUDA_VISIBLE_DEVICES=7 python inference_1.3b.py --pose_video I_GO_SCHOOL_LATER_20251215_142204_openpose_style.mp4 --ref_image square8.png