From afd1d92722c4499dc56887d1f8b5184160ec06c0 Mon Sep 17 00:00:00 2001
From: Jaret Burkett <jaretburkett@gmail.com>
Date: Mon, 17 Aug 2026 15:39:07 -0600
Subject: [PATCH] Fix issue with double sampling qwen vl video frames for
 hidream h3 for reference videos

---
 .../diffusion_models/minimax_h3/src/text_encoder.py      | 9 ++++++++-
 toolkit/dataloader_mixins.py                             | 5 +++++
 version.py                                               | 2 +-
 3 files changed, 14 insertions(+), 2 deletions(-)

diff --git a/extensions_built_in/diffusion_models/minimax_h3/src/text_encoder.py b/extensions_built_in/diffusion_models/minimax_h3/src/text_encoder.py
index 8a7f103..f58863d 100644
--- a/extensions_built_in/diffusion_models/minimax_h3/src/text_encoder.py
+++ b/extensions_built_in/diffusion_models/minimax_h3/src/text_encoder.py
@@ -160,8 +160,15 @@ def encode_minimax_h3_prompt(
             pixel_values = vision["pixel_values"]
             image_grid_thw = vision["image_grid_thw"]
         if videos:
+            import numpy as np
+
+            # frames are already sampled at 2 fps: do_sample_frames=False keeps
+            # the video processor from resampling them (official diffusers
+            # ref2va encoder passes the same flag)
             vids = processor.video_processor(
-                videos=[v.frames for v in videos], return_tensors="pt"
+                videos=[np.stack([np.asarray(f.convert("RGB")) for f in v.frames]) for v in videos],
+                do_sample_frames=False,
+                return_tensors="pt",
             )
             pixel_values_videos = vids["pixel_values_videos"]
             video_grid_thw = vids["video_grid_thw"]
diff --git a/toolkit/dataloader_mixins.py b/toolkit/dataloader_mixins.py
index bd460ec..59e592e 100644
--- a/toolkit/dataloader_mixins.py
+++ b/toolkit/dataloader_mixins.py
@@ -2136,6 +2136,11 @@ class TextEmbeddingFileItemDTOMixin:
             item["control_path"] = self.control_path
         if self.encode_control_in_text_embeddings and getattr(self, 'control_video_paths', None):
             item["control_videos"] = sorted(self.control_video_paths)
+            # v2: reference-video vision blocks are no longer resampled by the
+            # processor (do_sample_frames=False); older video-ref embeds are
+            # misaligned with their presentation. Only items WITH control
+            # videos carry this key, so no other cache is touched.
+            item["control_videos_version"] = 2
         # first-frame vision conditioning changes the embedding content -> new cache key
         elif (
             getattr(self, "encode_first_frame_in_text_embeddings", False)
diff --git a/version.py b/version.py
index 3ccd788..030fc75 100644
--- a/version.py
+++ b/version.py
@@ -1 +1 @@
-VERSION = "0.12.24"
+VERSION = "0.12.25"