Layer 3: 延迟 decode，只解码 CLIP 需要的 1 帧

- world model 调用 decode_video=False，跳过 16 帧全量 decode
- 只 decode 最后 1 帧给 CLIP embedding / observation queue
- 存 raw latent，循环结束后统一 batch decode 生成最终视频
- 每轮省 15 次 VAE decode，8 轮共省 120 次
- 跳过中间迭代的 wm tensorboard/mp4 保存

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

scripts/evaluation/world_model_interaction.py

@@ -559,6 +559,7 @@ def image_guided_synthesis_sim_mode(
         autocast_ctx = nullcontext()
 
     batch_variants = None
+    samples = None
     if ddim_sampler is not None:
         with autocast_ctx:
             samples, actions, states, intermedia = ddim_sampler.sample(
@@ -583,7 +584,7 @@ def image_guided_synthesis_sim_mode(
             batch_images = model.decode_first_stage(samples)
             batch_variants = batch_images
 
-    return batch_variants, actions, states
+    return batch_variants, actions, states, samples
 
 
 def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
@@ -693,7 +694,7 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
             sample_save_dir = f'{video_save_dir}/wm/{fs}'
             os.makedirs(sample_save_dir, exist_ok=True)
             # For collecting interaction videos
-            wm_video = []
+            wm_latent = []
             # Initialize observation queues
             cond_obs_queues = {
                 "observation.images.top":
@@ -749,7 +750,7 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
 
                 # Use world-model in policy to generate action
                 print(f'>>> Step {itr}: generating actions ...')
-                pred_videos_0, pred_actions, _ = image_guided_synthesis_sim_mode(
+                pred_videos_0, pred_actions, _, _ = image_guided_synthesis_sim_mode(
                     model,
                     sample['instruction'],
                     observation,
@@ -791,7 +792,7 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
 
                 # Interaction with the world-model
                 print(f'>>> Step {itr}: interacting with world model ...')
-                pred_videos_1, _, pred_states = image_guided_synthesis_sim_mode(
+                pred_videos_1, _, pred_states, wm_samples = image_guided_synthesis_sim_mode(
                     model,
                     "",
                     observation,
@@ -804,12 +805,16 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
                     fs=model_input_fs,
                     text_input=False,
                     timestep_spacing=args.timestep_spacing,
-                    guidance_rescale=args.guidance_rescale)
+                    guidance_rescale=args.guidance_rescale,
+                    decode_video=False)
+
+                # Decode only the last frame for CLIP embedding in next iteration
+                last_frame_pixel = model.decode_first_stage(wm_samples[:, :, -1:, :, :])
 
                 for idx in range(args.exe_steps):
                     observation = {
                         'observation.images.top':
-                        pred_videos_1[0][:, idx:idx + 1].permute(1, 0, 2, 3),
+                        last_frame_pixel[0, :, 0:1].permute(1, 0, 2, 3),
                         'observation.state':
                         torch.zeros_like(pred_states[0][idx:idx + 1]) if
                         args.zero_pred_state else pred_states[0][idx:idx + 1],
@@ -827,30 +832,14 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
                                        pred_videos_0,
                                        sample_tag,
                                        fps=args.save_fps)
-                # Save videos environment changes via world-model interaction
-                sample_tag = f"{args.dataset}-vid{sample['videoid']}-wd-fs-{fs}/itr-{itr}"
-                log_to_tensorboard(writer,
-                                   pred_videos_1,
-                                   sample_tag,
-                                   fps=args.save_fps)
-
-                # Save the imagen videos for decision-making
-                if pred_videos_0 is not None:
-                    sample_video_file = f'{video_save_dir}/dm/{fs}/itr-{itr}.mp4'
-                    save_results(pred_videos_0.cpu(),
-                                 sample_video_file,
-                                 fps=args.save_fps)
-                # Save videos environment changes via world-model interaction
-                sample_video_file = f'{video_save_dir}/wm/{fs}/itr-{itr}.mp4'
-                save_results(pred_videos_1.cpu(),
-                             sample_video_file,
-                             fps=args.save_fps)
 
                 print('>' * 24)
-                # Collect the result of world-model interactions
+                # Store raw latent for deferred decode
-                wm_video.append(pred_videos_1[:, :, :args.exe_steps].cpu())
+                wm_latent.append(wm_samples[:, :, :args.exe_steps].cpu())
 
-            full_video = torch.cat(wm_video, dim=2)
+            # Deferred decode: batch decode all stored latents
+            full_latent = torch.cat(wm_latent, dim=2).to(device)
+            full_video = model.decode_first_stage(full_latent).cpu()
             sample_tag = f"{args.dataset}-vid{sample['videoid']}-wd-fs-{fs}/full"
             log_to_tensorboard(writer,
                                full_video,