Update app.py

app.py

@@ -254,6 +254,13 @@ def _convert_diffusion_model_lora(raw: dict, base_shapes: dict, swap_fc1: bool)
     # we never fetch, not as tensors), so these fall to the same `alpha = rank` default every other module gets.
     native_ab = re.compile(r"^(.+)\.(lora_[AB])\.\w+\.weight$")
 
+    # Family D: already-diffusers-native base names, but with Kohya's `lora_down`/`lora_up` suffixes and no
+    # `.weight` — confirmed against `lora1`'s own debug dump (`transformer_blocks.0.attn.to_k.lora_down`,
+    # `transformer_blocks.0.adaln_proj.linear.lora_down`). Distinct from `native_ab` above (needs an adapter-name
+    # segment plus `.weight`) and from `kohya_ab` below (needs the fused `lora_unet_blocks_N_TARGET` prefix). No
+    # qkv-splitting needed either way: these base names already have separate `to_q`/`to_k`/`to_v`.
+    native_kohya_suffix_ab = re.compile(r"^(?:diffusion_model\.)?(.+)\.(lora_down|lora_up)$")
+
     # Family B (Kohya-style): `lora_unet_blocks_N_TARGET.(lora_down|lora_up|alpha)` — covers SB and Fluid
     # Enhancer. `lora_down`/`lora_up` are the same A/B convention under a different name.
     kohya_ab = re.compile(
@@ -275,6 +282,12 @@ def _convert_diffusion_model_lora(raw: dict, base_shapes: dict, swap_fc1: bool)
         elif name.startswith("blocks."):
             name = "transformer_blocks." + name[len("blocks."):]
         name = name.replace("final_layer.adaln_proj.linear", "norm_out.linear")
+        name = name.replace("final_layer.video_out", "proj_out")
+        name = name.replace("final_layer.audio_out", "audio_proj_out")
+        name = name.replace("final_layer.norm", "norm_out.norm")
+        name = name.replace("video_patch_proj", "proj_in")
+        name = name.replace("audio_patch_proj", "audio_proj_in")
+        name = name.replace("condition_proj", "context_embedder")
         name = name.replace(".attn.out_proj", ".attn.to_out.0")
         name = name.replace(".mlp.fc2", ".ff.net.2")
         name = name.replace(".mlp.fc1", ".ff.net.0.proj")
@@ -336,6 +349,17 @@ def _convert_diffusion_model_lora(raw: dict, base_shapes: dict, swap_fc1: bool)
                 print(f"[lora-convert] '{module_base}' isn't a real target — skipping", flush=True)
             continue
 
+        match = native_kohya_suffix_ab.match(key)
+        if match:
+            raw_base, direction = match.groups()
+            module_base = rename_base(raw_base)  # covers the final_layer.*/patch_proj/condition_proj renames too
+            ab = "lora_A" if direction == "lora_down" else "lora_B"
+            if f"{module_base}.weight" in base_shapes:
+                out[f"{LORA_KEY_PREFIX}.{module_base}.{ab}.weight"] = tensor
+            else:
+                print(f"[lora-convert] '{raw_base}' renamed to '{module_base}', which isn't a real target — skipping", flush=True)
+            continue
+
         match = standard_ab.match(key)
         if match:
             raw_base, ab = match.groups()
@@ -1417,6 +1441,7 @@ with gr.Blocks(title="PlagueKind · MiniMax-H3") as demo:
                     value=DEFAULT_INTERPOLATION,
                     info="MiniMax-H3 generates 24 fps; FILM synthesizes the frames in between.",
                 )
+                seed = gr.Number(label="Seed", value=42, precision=0)
                 upsample = gr.Checkbox(
                     label="Upsample prompt",
                     value=False,