diff --git a/movie_editor/backend/nodes.py b/movie_editor/backend/nodes.py
index d51addf6..93d187e5 100644
--- a/movie_editor/backend/nodes.py
+++ b/movie_editor/backend/nodes.py
@@ -194,7 +194,9 @@ ROLES: dict[str, dict] = {
 # Each entry: id, type, label, required, role_hint (role key to suggest adding), hint.
 # MiniMax H3 replaces two of these: its own EmptyMiniMaxH3LatentAV emits the video AND
 # audio streams in one node, so there is no separate audio-latent step and the latent is
-# no longer optional (nothing else produces one). The audio VAE stays required — it decodes
+# no longer optional — Studio does not generate one for this family, so a node has to. It
+# needn't be the Empty node: fl2va / ref2va build their own AV latent alongside their
+# conditioning, and satisfy this the same way. The audio VAE stays required — it decodes
 # the generated audio — and gains a second, optional job on H3 (encoding audio references).
 FAMILY_REQUIREMENTS: dict[str, dict] = {
     "minimax_h3": {
@@ -206,7 +208,9 @@ FAMILY_REQUIREMENTS: dict[str, dict] = {
             "clip": {"hint": "Add the MiniMax H3 text encoder (Qwen3-VL-32B, truncated to 50 layers)."},
             "init_latent": {"required": True, "label": "AV latent",
                             "hint": "Add Empty MiniMax H3 AV Latent — it makes the video and audio "
-                                    "streams together and feeds the Chain Sampler directly."},
+                                    "streams together and feeds the Chain Sampler directly. A "
+                                    "MiniMax H3 Image to Video (fl2va) or Reference to Video "
+                                    "(ref2va) node covers this too: each builds its own AV latent."},
         },
     },
 }
diff --git a/movie_editor/frontend/models.js b/movie_editor/frontend/models.js
index 26501de6..27c02951 100644
--- a/movie_editor/frontend/models.js
+++ b/movie_editor/frontend/models.js
@@ -1007,9 +1007,23 @@
     });
   }
 
-  // For VAE: need two separate producers (video_vae + audio_vae roles).
-  // For others: at least one producer of the type (regardless of role).
+  // Does this slot SOURCE the type — output it without consuming one? A loader or a
+  // generator does; a patcher that only passes the type through (a LoRA on MODEL/CLIP)
+  // does not, and can't satisfy a requirement on its own.
+  function slotSourcesType(type, slot) {
+    const spec = specFor(slot);
+    if (!(spec?.outputs || []).some((o) => o.type === type)) return false;
+    return !(spec?.connection_inputs || []).some((ci) => ci.type === type);
+  }
+
+  // VAE is asked for twice — video and audio — and only the role tells the two apart, so
+  // those stay role-bound (two separate producers required). Every other requirement is
+  // about whether the pipeline can GET that type from somewhere, so any node that really
+  // sources it counts, whatever role the user filed it under: MiniMax H3's fl2va / ref2va
+  // nodes emit their own AV latent, which is exactly what an Empty AV Latent would provide.
   function requirementSatisfied(req) {
+    if (req.type === "VAE") return slotProducesType(req.type, req.role_hint);
+    if (config.slots.some((s) => slotSourcesType(req.type, s))) return true;
     if (req.role_hint) return slotProducesType(req.type, req.role_hint);
     return slotProducesType(req.type, null);
   }