From 699bad38873bf8a772fbbcd1aad9234c1ee3286b Mon Sep 17 00:00:00 2001
From: kijai <40791699+kijai@users.noreply.github.com>
Date: Sat, 13 Sep 2025 20:10:13 +0300
Subject: [PATCH] InfiniteTalk: Don't stack or return frames when using output
 path

---
 HuMo/nodes.py      |  2 +-
 multitalk/nodes.py |  2 +-
 nodes.py           | 18 ++++++++++++------
 nodes_utility.py   |  7 ++++---
 4 files changed, 18 insertions(+), 11 deletions(-)

diff --git a/HuMo/nodes.py b/HuMo/nodes.py
index 507c554..f93e740 100644
--- a/HuMo/nodes.py
+++ b/HuMo/nodes.py
@@ -212,7 +212,7 @@ class HuMoEmbeds:
                 reference_images_in = common_upscale(reference_images.movedim(-1, 1), width, height, "lanczos", "disabled").movedim(-1, 1)
             else:
                 reference_images_in = reference_images
-            samples, = WanVideoEncodeLatentBatch.encode(self, vae, reference_images_in, tiled_vae, 0, 0, 0, 0)
+            samples, = WanVideoEncodeLatentBatch.encode(self, vae, reference_images_in, tiled_vae, None, None, None, None)
             samples = samples["samples"].transpose(0, 2).squeeze(0)
             num_refs = samples.shape[1]
 
diff --git a/multitalk/nodes.py b/multitalk/nodes.py
index b305237..4e5690b 100644
--- a/multitalk/nodes.py
+++ b/multitalk/nodes.py
@@ -391,7 +391,7 @@ class WanVideoImageToVideoMultiTalk:
                     "multitalk",
                     "infinitetalk"
                 ], {"default": "auto", "tooltip": "The sampling strategy to use in the long video generation loop, should match the model used"}),
-                "output_path": ("STRING", {"default": "", "tooltip": "If set, will save each window's resulting frames to this folder"}),
+                "output_path": ("STRING", {"default": "", "tooltip": "If set, will save each window's resulting frames to this folder, also DISABLES returning the final video tensor to save memory"}),
 
             }
         }
diff --git a/nodes.py b/nodes.py
index b7d0c00..4d030d9 100644
--- a/nodes.py
+++ b/nodes.py
@@ -3796,8 +3796,8 @@ class WanVideoSampler:
                                     im = Image.fromarray(video_np[i])
                                     im.save(os.path.join(output_path, f"frame_{img_counter:05d}.png"))
                                     img_counter += 1
-                                
-                            gen_video_list.append(videos if is_first_clip else videos[:, cur_motion_frames_num:])
+                            else:
+                                gen_video_list.append(videos if is_first_clip else videos[:, cur_motion_frames_num:])
 
                             current_condframe_index += 1
                             iteration_count += 1
@@ -3845,8 +3845,11 @@ class WanVideoSampler:
                                     last_frame = original_images[:, :, -1:, :, :]
                                     miss_length   = 1
                                     original_images = torch.cat([original_images, last_frame.repeat(1, 1, miss_length, 1, 1)], dim=2)
-
-                        gen_video_samples = torch.cat(gen_video_list, dim=1)
+                        
+                        if not output_path:
+                            gen_video_samples = torch.cat(gen_video_list, dim=1)
+                        else:
+                            gen_video_samples = torch.zeros(3, 1, 64, 64) # dummy output
 
                         if force_offload:
                             if not model["auto_cpu_offload"]:
@@ -3856,7 +3859,7 @@ class WanVideoSampler:
                             torch.cuda.reset_peak_memory_stats(device)
                         except:
                             pass
-                        return {"video": gen_video_samples.permute(1, 2, 3, 0)},
+                        return {"video": gen_video_samples.permute(1, 2, 3, 0), "output_path": output_path},
                     # region framepack loop
                     elif framepack:
                         framepack_out = []
@@ -4283,7 +4286,10 @@ class WanVideoEncodeLatentBatch:
 
         latent_list = []
         for img in images:
-            latent = vae.encode(img.unsqueeze(0).unsqueeze(0).permute(0, 4, 1, 2, 3), device=device, tiled=enable_vae_tiling, tile_size=(tile_x//vae.upsampling_factor, tile_y//vae.upsampling_factor), tile_stride=(tile_stride_x//vae.upsampling_factor, tile_stride_y//vae.upsampling_factor))
+            if enable_vae_tiling and tile_x is not None:
+                latent = vae.encode(img.unsqueeze(0).unsqueeze(0).permute(0, 4, 1, 2, 3), device=device, tiled=enable_vae_tiling, tile_size=(tile_x//vae.upsampling_factor, tile_y//vae.upsampling_factor), tile_stride=(tile_stride_x//vae.upsampling_factor, tile_stride_y//vae.upsampling_factor))
+            else:
+                latent = vae.encode(img.unsqueeze(0).unsqueeze(0).permute(0, 4, 1, 2, 3), device=device, tiled=enable_vae_tiling)
             vae.model.clear_cache()
             if latent_strength != 1.0:
                 latent *= latent_strength
diff --git a/nodes_utility.py b/nodes_utility.py
index bbb195d..fa84e28 100644
--- a/nodes_utility.py
+++ b/nodes_utility.py
@@ -456,8 +456,9 @@ class WanVideoPassImagesFromSamples:
                 }
                 }
 
-    RETURN_TYPES = ("IMAGE",)
-    RETURN_NAMES = ("images",)
+    RETURN_TYPES = ("IMAGE", "STRING",)
+    RETURN_NAMES = ("images", "output_path",)
+    OUTPUT_TOOLTIPS = ("Decoded images from the samples dictionary", "Output path if provided in the samples dictionary",)
     FUNCTION = "decode"
     CATEGORY = "WanVideoWrapper"
     DESCRIPTION = "Gets possible already decoded images from the samples dictionary, used with Multi/InfiniteTalk sampling"
@@ -466,7 +467,7 @@ class WanVideoPassImagesFromSamples:
         video = samples.get("video", None)
         video.clamp_(-1.0, 1.0)
         video.add_(1.0).div_(2.0)
-        return video.cpu().float(),
+        return video.cpu().float(), samples.get("output_path", "")
     
 NODE_CLASS_MAPPINGS = {
     "WanVideoImageResizeToClosest": WanVideoImageResizeToClosest,