Add temporal chunking to latent upscale

H3_LONG_VIDEOS_GUIDE.md

@@ -1165,7 +1165,7 @@ What this really means:
 - the node then runs a short refinement pass over the upscaled latent with the sampler, scheduler, step count, denoise, and megapixel target you picked on the latent-upscale params node
 - if the target is larger than the spatial tile size, that refinement pass is processed in spatial batches using the same tile defaults as the upstream latent-split node
 - the spatial stitch mode follows the upstream overlap controls, including `linear`, `smoothstep`, `overwrite`, and `midpoint`
-- the node also carries the upstream split compatibility knobs (`chunk_length`, `resize_conditioning`, and `anchor_strength`) so the control surface stays in one place
+- the node also carries the upstream split compatibility knobs (`chunk_length`, `temporal_overlap`, `resize_conditioning`, and `anchor_strength`) so the control surface stays in one place
 
 Good starting point:
 
@@ -1173,6 +1173,7 @@ Good starting point:
 - use the interpolation mode when you want a cheaper resize-only path
 - start with `euler_ancestral`, `simple`, `2` steps, and `0.2` denoise
 - leave width and height at `0` unless you want an exact override; otherwise `megapixels` drives the target size
+- for long shots on smaller cards, start with `chunk_length = 85` and `temporal_overlap = 17` so the latent upscaler works in shorter temporal passes
 - leave the spatial tile inputs at their defaults first: `512x512` tiles, `64` overlap, `0` fade width, `earlier` overlap mode
 - keep `linear` blend first unless you want to reproduce a specific upstream stitch style
 - leave the split compatibility knobs alone unless you specifically need to mirror the upstream node behavior

README.md

@@ -47,7 +47,7 @@
   - Per-shot directives now support `continuity:`, `ref_mode:`, `ref_noise_aug:`, `anchor_add:`, `soundscape:`, and `music:` in addition to the existing timing and wardrobe directives.
 
 - `Dumas H3 Latent Upscale Params`
-  - Inputs: `mode`, `model_name`, `method`, `width`, `height`, `device`, `precision`, `sampler_name`, `scheduler`, `steps`, `denoise`, `megapixels`, `tile_width`, `tile_height`, `overlap`, `fade_width`, `fade_height`, `overlap_mode`, `overlap_blend`, `tile_size_mode`, `grid_rows`, `grid_cols`, `spatial_w_overlap`, `spatial_h_overlap`, `min_tile_size`, `masked_area_noise`, `brightness_match`, `dynamic_fade`, `dynamic_fade_min`, `chunk_length`, `resize_conditioning`, `anchor_strength`
+  - Inputs: `mode`, `model_name`, `method`, `width`, `height`, `device`, `precision`, `sampler_name`, `scheduler`, `steps`, `denoise`, `megapixels`, `tile_width`, `tile_height`, `overlap`, `fade_width`, `fade_height`, `overlap_mode`, `overlap_blend`, `tile_size_mode`, `grid_rows`, `grid_cols`, `spatial_w_overlap`, `spatial_h_overlap`, `min_tile_size`, `masked_area_noise`, `brightness_match`, `dynamic_fade`, `dynamic_fade_min`, `chunk_length`, `temporal_overlap`, `resize_conditioning`, `anchor_strength`
   - Output: `latent_upscale_param`
   - Bundles the optional latent-space upscaler settings used by `Dumas H3 Long Videos` before decode, so the main node can rebuild conditioning at the target size and run a short refinement pass with your chosen sampler, scheduler, step count, denoise, and the full upstream spatial split controls.
 

dumas_h3_latent_upscale.py

@@ -372,6 +372,43 @@ def _resolve_target_size(param, h_in, w_in):
     return int(h_in), int(w_in)
 
 
+def _temporal_segments(frame_count, chunk_length, overlap):
+    frame_count = int(frame_count)
+    chunk_length = int(chunk_length)
+    overlap = int(overlap)
+    if frame_count <= 0:
+        return [(0, 0)]
+    if chunk_length <= 0:
+        raise ValueError("chunk_length must be positive")
+    if overlap < 0:
+        raise ValueError("temporal_overlap must be non-negative")
+    if chunk_length <= overlap:
+        raise ValueError("temporal_overlap must be smaller than chunk_length")
+    if frame_count <= chunk_length:
+        return [(0, frame_count)]
+
+    hop = chunk_length - overlap
+    bounds = []
+    start = 0
+    while start < frame_count:
+        end = min(start + chunk_length, frame_count)
+        bounds.append((start, end))
+        if end >= frame_count:
+            break
+        start += hop
+    return bounds
+
+
+def _temporal_blend_weights(length, anchor_strength):
+    length = int(length)
+    if length <= 0:
+        return None
+    anchor_strength = float(anchor_strength)
+    anchor_strength = min(1.0, max(0.0, anchor_strength))
+    start = max(0.0, 1.0 - anchor_strength)
+    return torch.linspace(start, 1.0, length)
+
+
 def _spatial_blend_weights(t, overlap_mode, overlap_blend="linear"):
     if overlap_blend == "overwrite":
         return torch.ones_like(t) if overlap_mode == "later" else torch.zeros_like(t)
@@ -633,13 +670,55 @@ def upscale_video_interp(video, param):
     return up, h_out, w_out
 
 
+def _upscale_video_temporal_chunks(video, param, upscaler):
+    chunk_length = int(param.get("chunk_length", 0) or 0)
+    temporal_overlap = int(param.get("temporal_overlap", 0) or 0)
+    anchor_strength = float(param.get("anchor_strength", 0.999) or 0.999)
+    t = int(video.shape[2])
+    if chunk_length <= 0 or t <= chunk_length:
+        return upscaler(video, param)
+
+    bounds = _temporal_segments(t, chunk_length, temporal_overlap)
+    if len(bounds) <= 1:
+        return upscaler(video, param)
+
+    orig_dtype = video.dtype
+    out = None
+    out_h = out_w = None
+    for i, (t0, t1) in enumerate(bounds):
+        chunk = video[:, :, t0:t1].contiguous()
+        chunk_out, chunk_h, chunk_w = upscaler(chunk, param)
+        chunk_out = chunk_out.to(device="cpu", dtype=orig_dtype)
+        if out is None:
+            out_h, out_w = chunk_h, chunk_w
+            out = torch.zeros((video.shape[0], video.shape[1], t, out_h, out_w), device="cpu", dtype=orig_dtype)
+            out[:, :, t0:t1] = chunk_out
+            continue
+
+        ov = min(temporal_overlap, t1 - t0, t - t0)
+        if ov <= 0:
+            out[:, :, t0:t1] = chunk_out
+            continue
+
+        prev_region = out[:, :, t0:t0 + ov]
+        new_region = chunk_out[:, :, :ov]
+        w = _temporal_blend_weights(ov, anchor_strength).to(device=prev_region.device, dtype=prev_region.dtype)
+        out[:, :, t0:t0 + ov] = (
+            prev_region * (1.0 - w[None, None, :, None, None]) +
+            new_region * w[None, None, :, None, None]
+        )
+        out[:, :, t0 + ov:t1] = chunk_out[:, :, ov:]
+
+    return out, out_h, out_w
+
+
 def upscale_latent_video(video, param):
     mode = str(param.get("mode") or "off")
     if mode == "off":
         return video, video.shape[-2], video.shape[-1]
     if mode == "model":
-        return upscale_video_model(video, param)
-    return upscale_video_interp(video, param)
+        return _upscale_video_temporal_chunks(video, param, upscale_video_model)
+    return _upscale_video_temporal_chunks(video, param, upscale_video_interp)
 
 
 class H3LatentUpscaleParams:
@@ -711,8 +790,10 @@ class H3LatentUpscaleParams:
                     "tooltip": "Temporal fade schedule over each tile's sampling. Off keeps the fade fixed; narrowing shrinks it over steps; widening grows it over steps."}),
                 "dynamic_fade_min": ("INT", {"default": 32, "min": 0, "max": 4096, "step": 32,
                     "tooltip": "Minimum fade width used by dynamic_fade when it is enabled."}),
-                "chunk_length": ("INT", {"default": 136, "min": 17, "max": 100000, "step": 17,
-                    "tooltip": "Reserved for upstream split compatibility. 136 matches the original chunk-length default."}),
+                "chunk_length": ("INT", {"default": 85, "min": 17, "max": 100000, "step": 17,
+                    "tooltip": "Temporal chunk length for the latent upscale stage. 85 matches the practical upstream example and helps lower peak VRAM."}),
+                "temporal_overlap": ("INT", {"default": 17, "min": 0, "max": 100000, "step": 17,
+                    "tooltip": "Temporal overlap between latent chunks. 17 matches the upstream split example and reduces seam risk."}),
                 "resize_conditioning": ("BOOLEAN", {"default": False,
                     "tooltip": "Reserved for upstream split compatibility. Leave OFF unless you need the original fallback behavior."}),
                 "anchor_strength": ("FLOAT", {"default": 0.999, "min": 0.0, "max": 1.0, "step": 0.01,
@@ -720,7 +801,7 @@ class H3LatentUpscaleParams:
             }
         }
 
-    def build(self, mode, model_name, method, width, height, device, precision, sampler_name, scheduler, steps, denoise, megapixels, tile_width, tile_height, overlap, fade_width, fade_height, overlap_mode, overlap_blend, tile_size_mode, grid_rows, grid_cols, spatial_w_overlap, spatial_h_overlap, min_tile_size, masked_area_noise, brightness_match, dynamic_fade, dynamic_fade_min, chunk_length, resize_conditioning, anchor_strength):
+    def build(self, mode, model_name, method, width, height, device, precision, sampler_name, scheduler, steps, denoise, megapixels, tile_width, tile_height, overlap, fade_width, fade_height, overlap_mode, overlap_blend, tile_size_mode, grid_rows, grid_cols, spatial_w_overlap, spatial_h_overlap, min_tile_size, masked_area_noise, brightness_match, dynamic_fade, dynamic_fade_min, chunk_length, temporal_overlap, resize_conditioning, anchor_strength):
         width = int(width)
         height = int(height)
         steps = int(steps)
@@ -737,6 +818,7 @@ class H3LatentUpscaleParams:
         masked_area_noise = float(masked_area_noise)
         dynamic_fade_min = int(dynamic_fade_min)
         chunk_length = int(chunk_length)
+        temporal_overlap = int(temporal_overlap)
         anchor_strength = float(anchor_strength)
         if mode == "off":
             return ({
@@ -767,9 +849,16 @@ class H3LatentUpscaleParams:
                 "dynamic_fade": dynamic_fade,
                 "dynamic_fade_min": dynamic_fade_min,
                 "chunk_length": chunk_length,
+                "temporal_overlap": temporal_overlap,
                 "resize_conditioning": bool(resize_conditioning),
                 "anchor_strength": anchor_strength,
             },)
+        if chunk_length % 17 != 0:
+            raise ValueError("chunk_length must be a multiple of 17 pixels")
+        if temporal_overlap % 17 != 0:
+            raise ValueError("temporal_overlap must be a multiple of 17 pixels")
+        if temporal_overlap >= chunk_length:
+            raise ValueError("temporal_overlap must be smaller than chunk_length")
         if width > 0:
             width = int(round(width / 32.0)) * 32
         if height > 0:
@@ -806,6 +895,7 @@ class H3LatentUpscaleParams:
             "dynamic_fade": dynamic_fade,
             "dynamic_fade_min": dynamic_fade_min,
             "chunk_length": chunk_length,
+            "temporal_overlap": temporal_overlap,
             "resize_conditioning": bool(resize_conditioning),
             "anchor_strength": anchor_strength,
         },)

tests/test_dumas_h3_longvideos.py

@@ -1144,7 +1144,8 @@ class DumasH3LongVideosHelperTests(unittest.TestCase):
         self.assertFalse(required["brightness_match"][1]["default"])
         self.assertEqual(required["dynamic_fade"][1]["default"], "off")
         self.assertEqual(required["dynamic_fade_min"][1]["default"], 32)
-        self.assertEqual(required["chunk_length"][1]["default"], 136)
+        self.assertEqual(required["chunk_length"][1]["default"], 85)
+        self.assertEqual(required["temporal_overlap"][1]["default"], 17)
         self.assertFalse(required["resize_conditioning"][1]["default"])
         self.assertEqual(required["anchor_strength"][1]["default"], 0.999)
 
@@ -1175,6 +1176,13 @@ class DumasH3LongVideosHelperTests(unittest.TestCase):
         self.assertLess(smaller["tile_height"], 512)
         self.assertEqual(smaller["fade_width"] % 32, 0)
         self.assertEqual(smaller["fade_height"] % 32, 0)
+        self.assertLessEqual(smaller["min_tile_size"], smaller["tile_width"])
+        self.assertLessEqual(smaller["min_tile_size"], smaller["tile_height"])
+
+    def test_temporal_segments_split_long_sequences(self):
+        latent = importlib.import_module("dumas_h3_latent_upscale")
+        bounds = latent._temporal_segments(124, 85, 17)
+        self.assertEqual(bounds, [(0, 85), (68, 124)])
 
     def test_compose_persistent_does_not_expand_ambiguous_plural_to_full_cast(self):
         active = self.module.parse_wardrobe(