Adjust MiniMah h3 sizing for videos to downscale to match

extensions_built_in/diffusion_models/minimax_h3/minimax_h3.py

@@ -1330,7 +1330,7 @@ class MinimaxH3Ref2VAModel(MinimaxH3Model):
         blocks = [(1, h, w, 0) for h, w in ref_shapes]
         audio_rows = []
         self._append_video_ref_blocks(
-            batch, all_rows, audio_rows, blocks, device, dtype
+            batch, all_rows, audio_rows, blocks, device, dtype, target_h, target_w
         )
         if not all_rows:
             return None, None, (), ()
@@ -1364,8 +1364,10 @@ class MinimaxH3Ref2VAModel(MinimaxH3Model):
             n = packing.align_num_frames_down(max(len(frames), 5))
             frames = frames[:n]
         h0, w0 = frames[0].shape[:2]
-        # ComfyUI reference-video sizing (independent of the sample canvas)
-        ph, pw = packing.reference_video_pixel_size(w0, h0)
+        # match the sample canvas's pixel area, own aspect kept
+        ph, pw = packing.reference_video_pixel_size(
+            w0, h0, gen_config.height, gen_config.width
+        )
         pixels = torch.from_numpy(np.stack(frames)).float() / 255.0 * 2.0 - 1.0
         pixels = pixels.permute(3, 0, 1, 2)[None]  # (1, 3, T, H, W)
         pixels = (
@@ -1405,10 +1407,11 @@ class MinimaxH3Ref2VAModel(MinimaxH3Model):
         return {"latent": latents[0].float(), "audio_rows": audio_rows}
 
     def _append_video_ref_blocks(
-        self, batch, all_rows, audio_rows, blocks, device, dtype
+        self, batch, all_rows, audio_rows, blocks, device, dtype, target_h, target_w
     ):
-        # control videos get dataset-identical treatment (frame count, fps,
-        # bucket) and one VAE encode, disk-cached next to the video; the
+        # control videos get the dataset's temporal treatment (frame count,
+        # fps), are area-matched to the target, and get one VAE encode
+        # disk-cached next to the video; the
         # resulting latents become multi-frame reference blocks packed after
         # the image references
         paths_per_item = getattr(batch, "control_video_paths_list", None)
@@ -1425,7 +1428,7 @@ class MinimaxH3Ref2VAModel(MinimaxH3Model):
             auds = []
             for per_item in paths_per_item:
                 entry = load_ref_video_latent(
-                    self, per_item[ref_idx], batch.dataset_config
+                    self, per_item[ref_idx], batch.dataset_config, target_h, target_w
                 )
                 lats.append(entry["latent"].to(device, torch.float32))
                 auds.append(entry.get("audio_rows"))

extensions_built_in/diffusion_models/minimax_h3/src/packing.py

@@ -124,16 +124,17 @@ def reference_pixel_size(
     return height, width
 
 
-def reference_video_pixel_size(ref_width: int, ref_height: int) -> Tuple[int, int]:
-    """Reference VIDEO sizing (ComfyUI): the 768-short-edge canvas with the
-    768*1344 area cap, or the native size (rounded to /32) when the source is
-    smaller than that canvas. Returns (height, width)."""
-    ch, cw = resolve_canvas_size(ref_width, ref_height)
-    if ref_width * ref_height < cw * ch:
-        m = CANVAS_MULTIPLE
-        cw = max(m, round(ref_width / m) * m)
-        ch = max(m, round(ref_height / m) * m)
-    return ch, cw
+def reference_video_pixel_size(
+    ref_width: int, ref_height: int, target_height: int, target_width: int
+) -> Tuple[int, int]:
+    """Reference VIDEO sizing: match the TARGET's pixel area with the ref's own
+    aspect kept (same aspect -> exactly the target size; other aspects -> the
+    same pixel budget on the /32 grid). Returns (height, width)."""
+    scale = math.sqrt((target_height * target_width) / float(ref_width * ref_height))
+    m = CANVAS_MULTIPLE
+    height = max(m, round(ref_height * scale / m) * m)
+    width = max(m, round(ref_width * scale / m) * m)
+    return height, width
 
 
 def prepare_reference_image(

extensions_built_in/diffusion_models/minimax_h3/src/ref_video_cache.py

@@ -1,8 +1,8 @@
 """Reference-video latents for ref2va, without dataloader machinery.
 
 A control VIDEO gets the dataset's temporal treatment — num_frames /
-auto_frame_count, fps — and ComfyUI's reference sizing (768-short-edge canvas
-or native when smaller, aspect kept), then a single VAE encode whose
+auto_frame_count, fps — and is area-matched to the target (own aspect kept, /32
+grid), then a single VAE encode whose
 result is cached next to the video in ``_latent_cache/`` (keyed like normal
 latent caches: file signature + the config values that shape the latent).
 Everything is deterministic (even frame spread, no random start) so the cache
@@ -112,7 +112,9 @@ def _cache_path(path: str, hash_dict: dict) -> str:
 
 
 @torch.no_grad()
-def load_ref_video_latent(model, path: str, dataset_config) -> dict:
+def load_ref_video_latent(
+    model, path: str, dataset_config, target_height: int, target_width: int
+) -> dict:
     """Returns {"latent": (C, T, h, w) cpu tensor, "num_frames": int},
     encoding + disk-caching on first use. ``model`` is the MinimaxH3 model
     (used for the VAE, audio encode and the frame-count snapper)."""
@@ -145,7 +147,10 @@ def load_ref_video_latent(model, path: str, dataset_config) -> dict:
     )
     hash_dict = {
         "signature": get_quick_signature_string(path),
-        "ref_sizing": "comfy_canvas",
+        "ref_sizing": "match_target_area",
+        "target_area": int(target_height * target_width)
+        if target_height and target_width
+        else 0,
         "num_frames": num_frames,
         "fps": dataset_config.fps,
         "trim_tail": trim_tail,
@@ -164,9 +169,10 @@ def load_ref_video_latent(model, path: str, dataset_config) -> dict:
         mem_cache[path] = entry
         return entry
 
-    # ComfyUI reference-video sizing: 768-short-edge canvas (768*1344 area
-    # cap) or native size when smaller; aspect-preserving resize, no crop
-    out_h, out_w = reference_video_pixel_size(src_w, src_h)
+    # match the target's pixel area (the dataset bucket the target trains at)
+    # with the ref's own aspect: same aspect -> identical size; aspect-
+    # preserving resize, no crop
+    out_h, out_w = reference_video_pixel_size(src_w, src_h, target_height, target_width)
 
     indices = ref_frame_indices(
         total, src_fps, num_frames, dataset_config.fps, trim_tail

ui/src/app/jobs/new/options.tsx

@@ -922,7 +922,7 @@ export const modelArchs: ModelArch[] = [
       <div className="space-y-2">
         <p>
           Reference-to-video: control images condition the output as subject/style references (never as a first frame).
-          Reference images keep their aspect and scale down (never up) to the target's pixel area; reference videos use the 768-short-edge canvas (or native size if smaller). Each rides into the packed
+          References keep their own aspect and are matched to the target's pixel area (images scale down only, never up; a same-aspect video reference is exactly the target size). Each rides into the packed
           sequence as a reference block, and is also shown to the Qwen3-VL conditioner as a{' '}
           <code>&lt;Picture i&gt;</code> vision block. Training references come from the dataset control path(s);
           sampling uses the sample ctrl images — always as references. Image references only for now (no reference

version.py

@@ -1 +1 @@
-VERSION = "0.12.22"
+VERSION = "0.12.23"