diff --git a/pyproject.toml b/pyproject.toml
index 0f57ecf4df..a89d2c2a51 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -88,7 +88,11 @@ huggingfacenotorch = [
     "peft>=0.18.0,!=0.11.0",
     "huggingface_hub>=0.34.0",
     "hf_transfer",
-    "diffusers",
+    # 0.39 is the floor the diffusion stack already assumes: Krea2Pipeline (the Krea 2
+    # workflow raises without it), the cache_context child-registry behavior, and the
+    # Flux2 / Z-Image pipelines. An unversioned entry let an upgrade keep an older
+    # diffusers, so advertised models failed to load until the user upgraded by hand.
+    "diffusers>=0.39.0",
     "transformers>=4.51.3,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3,!=4.53.0,!=4.54.0,!=4.55.0,!=4.55.1,!=4.57.0,!=4.57.4,!=4.57.5,!=5.0.0,!=5.1.0,<=5.5.0",
     "trl>=0.18.2,!=0.19.0,<=0.24.0",
     "sentence-transformers",
diff --git a/studio/backend/hub/services/models/deletion.py b/studio/backend/hub/services/models/deletion.py
index efc5013b61..8ef6f0e174 100644
--- a/studio/backend/hub/services/models/deletion.py
+++ b/studio/backend/hub/services/models/deletion.py
@@ -629,9 +629,14 @@ def _video_blocks_delete(repo_id: str) -> Optional[str]:
         logger.debug(f"Video backend unavailable during delete guard for {repo_id}: {e}")
         return None
     status = backend.status()
-    if status.get("loaded") and status.get("repo_id"):
-        if _loaded_id_matches_repo(str(status["repo_id"]), repo_id):
-            return "Unload the model before deleting"
+    if status.get("loaded"):
+        # repo_id names the checkpoint; for a GGUF / single-file load the companion base
+        # supplies the VAE and text encoders and is just as much part of the live model,
+        # so refuse it too (the Images guard above does the same via loaded_repo_ids).
+        for key in ("repo_id", "base_repo"):
+            held = status.get(key)
+            if held and _loaded_id_matches_repo(str(held), repo_id):
+                return "Unload the model before deleting"
     for lid in getattr(backend, "loading_repo_ids", tuple)():
         if _loaded_id_matches_repo(str(lid), repo_id):
             return "A Video model load is using this repo; wait for it to finish"
diff --git a/studio/backend/tests/test_video_routes.py b/studio/backend/tests/test_video_routes.py
index c127bc0cbe..3e7d0bb66a 100644
--- a/studio/backend/tests/test_video_routes.py
+++ b/studio/backend/tests/test_video_routes.py
@@ -734,3 +734,26 @@ def test_export_endpoint_validation(client, monkeypatch):
     resp = client.get(f"/api/inference/video/gallery/{video['id']}/export?format=webm")
     assert resp.status_code == 501
     assert "PyAV" in resp.json()["detail"]
+
+
+def test_delete_guard_protects_the_loaded_video_companion_base(monkeypatch):
+    # For a GGUF / single-file video load the companion base supplies the VAE and text
+    # encoders, so it is as much part of the live model as the checkpoint. Deleting it used
+    # to sail past the guard, which only compared repo_id.
+    from hub.services.models import deletion
+
+    class _Backend:
+        def status(self):
+            return {
+                "loaded": True,
+                "repo_id": "unsloth/LTX-2.3-GGUF",
+                "base_repo": "unsloth/LTX-2.3",
+            }
+
+        def loading_repo_ids(self):
+            return ()
+
+    monkeypatch.setattr(video_module, "get_video_backend", lambda: _Backend())
+    assert deletion._video_blocks_delete("unsloth/LTX-2.3-GGUF") is not None
+    assert deletion._video_blocks_delete("unsloth/LTX-2.3") is not None
+    assert deletion._video_blocks_delete("unsloth/something-else") is None
diff --git a/studio/frontend/src/features/model-picker/components/model-selector/pickers.tsx b/studio/frontend/src/features/model-picker/components/model-selector/pickers.tsx
index d13d84362c..9a5e972db3 100644
--- a/studio/frontend/src/features/model-picker/components/model-selector/pickers.tsx
+++ b/studio/frontend/src/features/model-picker/components/model-selector/pickers.tsx
@@ -1921,8 +1921,11 @@ export function HubModelPicker({
           ? isKnownGgufRepo(id)
           : !chatOnly || isRecommendableFormat(id, isKnownGgufRepo(id), isMac),
       )
-      // Member repos of a catalog group collapse into the canonical group row.
-      .filter((id) => !catalog || !groupForRepoId(id, catalog))
+      // Member repos of a catalog group would collapse into the canonical group row --
+      // but nothing renders those rows yet (catalogGroupFitsDevice / groupMatchesQuery are
+      // imported and unused), and a task-scoped picker's `models` is catalogToModelOptions(),
+      // i.e. group members exclusively. Suppressing them here emptied Recommended on the
+      // Images and Video pages, so keep the artifacts listed until the grouped UI lands.
       .filter((id) => !/-FP8[-.]|FP8-Dynamic/i.test(id));
     // Sort: GGUFs first, then hub models
     const gguf: string[] = [];
@@ -1970,8 +1973,8 @@ export function HubModelPicker({
         : rows.filter((r) => matchesFormatFilter(r.id, r.isGguf, formatFilter));
     // Task-scoped pages load single-file GGUF only.
     if (task) rows = rows.filter((r) => r.isGguf);
-    // Members already render under their canonical group row.
-    if (catalog) rows = rows.filter((r) => !groupForRepoId(r.id, catalog));
+    // Members would render under their canonical group row, but that row does not exist
+    // yet (see recommendedIds): filtering here removed curated models from Hub search too.
     // The "recommended" sort always applies the device-fit filter; the shared
     // "Fits on device" tick extends it to the other sorts too.
     if (recommendedSort !== "recommended" && !fitOnDeviceOnly) return rows;