Fix WanGP media routing and model controls

Author: georgi

Dockerfile.combined

@@ -5,7 +5,7 @@ ARG CORE_IMAGE=ghcr.io/nodetool-ai/nodetool:latest
 FROM ${CORE_IMAGE}
 
 ARG NODETOOL_CORE_REPOSITORY=https://github.com/nodetool-ai/nodetool-core.git
-ARG NODETOOL_CORE_COMMIT=a7938288e894af65764287e626ea8d0924bf954a
+ARG NODETOOL_CORE_COMMIT=2f5c2519d9e46b38975c349c8467e41e053ed0e8
 ARG WANGP_REPOSITORY=https://github.com/deepbeepmeep/Wan2GP.git
 ARG WANGP_COMMIT=362c3467a70e1136ceb52eec95907205a8f88543
 ARG MMGP_REPOSITORY=https://github.com/deepbeepmeep/mmgp.git
@@ -15,7 +15,7 @@ LABEL org.opencontainers.image.title="NodeTool WanGP non-commercial GPU worker"
       org.opencontainers.image.description="Authenticated direct-Python WanGP provider worker; free non-monetized evaluation only" \
       org.opencontainers.image.licenses="AGPL-3.0-or-later AND LicenseRef-WanGP-Community-2.0 AND LicenseRef-mmGP-NonCommercial" \
       org.opencontainers.image.source="https://github.com/nodetool-ai/nodetool-wan2gp" \
-      io.nodetool.core.commit="a7938288e894af65764287e626ea8d0924bf954a" \
+      io.nodetool.core.commit="2f5c2519d9e46b38975c349c8467e41e053ed0e8" \
       io.nodetool.wangp.commit="362c3467a70e1136ceb52eec95907205a8f88543" \
       io.nodetool.mmgp.commit="589ba050d320c4df879f48d11b23d6a88e6c68c8" \
       io.nodetool.commercial-use="prohibited-without-separate-written-license"
@@ -89,7 +89,7 @@ ENV WANGP_CONFIG_DIR=/workspace/wan2gp/config \
     WANGP_ROOT=/opt/Wan2GP \
     NODETOOL_MODEL_PREPARE_COMMAND_WANGP="/opt/wan2gp-venv/bin/python /opt/nodetool-wan2gp-combined/prepare_model.py" \
     NODETOOL_PROVIDER_ADAPTER_COMMAND_WANGP="/opt/wan2gp-venv/bin/python /opt/nodetool-wan2gp-combined/provider_adapter.py" \
-    NODETOOL_PROVIDER_ADAPTER_CAPABILITIES_WANGP="text_to_video,image_to_video,text_to_image,image_to_image,text_to_audio,text_to_speech" \
+    NODETOOL_PROVIDER_ADAPTER_CAPABILITIES_WANGP="text_to_video,image_to_video,text_to_image,image_to_image,text_to_audio,text_to_speech_encoded" \
     NODETOOL_PROVIDER_ADAPTER_DISPLAY_NAME_WANGP="WanGP" \
     HF_HOME=/workspace/cache/huggingface \
     PYTHONUNBUFFERED=1 \

combined/provider_adapter.py

@@ -93,6 +93,17 @@ def _is_music(metadata: dict[str, Any]) -> bool:
     return str(metadata.get("family") or "").casefold() == "music"
 
 
+def _choice_values(definition: Any) -> list[str]:
+    if not isinstance(definition, dict):
+        return []
+    values: list[str] = []
+    for choice in definition.get("choices", []):
+        value = choice.get("value") if isinstance(choice, dict) else choice
+        if value is not None and str(value):
+            values.append(str(value))
+    return values
+
+
 def _models(session: Any, model_kind: str) -> list[dict[str, Any]]:
     """Translate WanGP model metadata into NodeTool provider models."""
     if model_kind not in {"video", "image", "tts", "music"}:
@@ -127,7 +138,42 @@ def _models(session: Any, model_kind: str) -> list[dict[str, Any]]:
             "provider": "wangp",
         }
         if model_kind == "tts":
-            model["capabilities"] = supported
+            base_model_type = str(metadata.get("base_model_type") or model_type)
+            tts_capabilities = list(supported)
+            media_inputs = metadata.get("media_inputs")
+            audio_inputs = (
+                media_inputs.get("audio", {}) if isinstance(media_inputs, dict) else {}
+            )
+            if audio_inputs.get("prompt"):
+                tts_capabilities.append("voice_cloning")
+            if base_model_type in {"qwen3_tts_base", "omnivoice"}:
+                tts_capabilities.append("reference_transcript")
+            if base_model_type in {
+                "qwen3_tts_customvoice",
+                "index_tts2",
+                "index_tts25",
+            }:
+                tts_capabilities.append("instruction_control")
+            if base_model_type in {"qwen3_tts_voicedesign", "omnivoice"}:
+                tts_capabilities.append("voice_design")
+
+            setting_values = metadata.get("setting_values")
+            model_mode = (
+                setting_values.get("model_mode")
+                if isinstance(setting_values, dict)
+                else None
+            )
+            mode_label = str(
+                model_mode.get("label", "") if isinstance(model_mode, dict) else ""
+            ).casefold()
+            mode_values = _choice_values(model_mode)
+            if mode_label == "speaker":
+                tts_capabilities.append("preset_voice")
+                model["voices"] = mode_values
+            elif mode_label == "language":
+                tts_capabilities.append("language_selection")
+                model["languages"] = mode_values
+            model["capabilities"] = list(dict.fromkeys(tts_capabilities))
         else:
             model["supportedTasks"] = supported
         models.append(model)
@@ -205,6 +251,9 @@ def _settings(
     elif params.get("durationSeconds") is not None and operation.endswith("video"):
         settings["video_length"] = f"{float(params['durationSeconds']):g}s"
 
+    if operation in {"text_to_image", "image_to_image"}:
+        settings["image_mode"] = 1
+
     if operation in {"image_to_image", "image_to_video"}:
         image_path = str(_input_path(request, "image_path"))
         image_inputs = (metadata or {}).get("media_inputs", {}).get("image", {})
@@ -226,8 +275,12 @@ def _settings(
     if operation == "text_to_audio":
         style_prompt = str(params.get("prompt") or "")
         lyrics = str(params.get("lyrics") or "").strip()
-        settings["prompt"] = lyrics or "[Instrumental]"
-        settings["alt_prompt"] = style_prompt
+        base_model_type = str((metadata or {}).get("base_model_type") or model_type)
+        if base_model_type.startswith("stable_audio3"):
+            settings["prompt"] = style_prompt
+        else:
+            settings["prompt"] = lyrics or "[Instrumental]"
+            settings["alt_prompt"] = style_prompt
         if params.get("durationSeconds") is not None:
             settings["duration_seconds"] = float(params["durationSeconds"])
 
@@ -236,11 +289,16 @@ def _settings(
             settings["alt_prompt"] = str(params["referenceText"])
         elif params.get("instructions") is not None:
             settings["alt_prompt"] = str(params["instructions"])
-        model_mode = params.get("voice") or params.get("language")
+        base_model_type = str((metadata or {}).get("base_model_type") or model_type)
+        model_mode = (
+            params.get("voice")
+            if base_model_type == "qwen3_tts_customvoice"
+            else params.get("language")
+        )
         if model_mode:
             settings["model_mode"] = str(model_mode)
-        if params.get("speed") is not None:
-            settings["speech_speed"] = float(params["speed"])
+        if params.get("speed") is not None and base_model_type == "index_tts25":
+            settings["custom_settings"] = {"speech_speed": float(params["speed"])}
         if request.get("reference_audio_path"):
             settings["audio_guide"] = str(
                 _input_path(request, "reference_audio_path")
@@ -276,9 +334,18 @@ def _generated_path(result: Any, media_type: str | None = None) -> str:
         path = getattr(artifact, "path", None)
         if path and Path(path).is_file():
             return str(Path(path).resolve())
+    media_suffixes = {
+        "image": {".bmp", ".gif", ".jpeg", ".jpg", ".png", ".tif", ".tiff", ".webp"},
+        "video": {".avi", ".mkv", ".mov", ".mp4", ".webm"},
+        "audio": {".aac", ".flac", ".m4a", ".mp3", ".ogg", ".opus", ".wav"},
+    }
+    expected_suffixes = media_suffixes.get(media_type or "")
     for path in getattr(result, "generated_files", ()):
         candidate = Path(str(path))
-        if candidate.is_file():
+        if candidate.is_file() and (
+            expected_suffixes is None
+            or candidate.suffix.casefold() in expected_suffixes
+        ):
             return str(candidate.resolve())
     raise RuntimeError("WanGP completed without a generated media file")
 

tests/test_provider_adapter.py

@@ -73,6 +73,17 @@ def test_models_exposes_image_tts_and_music_separately() -> None:
                 "family": "tts",
                 "main_output": ["audio"],
                 "capabilities": {"text_to_audio": True},
+                "base_model_type": "qwen3_tts_base",
+                "media_inputs": {"audio": {"prompt": True}},
+                "setting_values": {
+                    "model_mode": {
+                        "label": "Language",
+                        "choices": [
+                            {"label": "English", "value": "English"},
+                            {"label": "Auto", "value": "auto"},
+                        ],
+                    }
+                },
             },
             {
                 "model_type": "ace_step",
@@ -97,7 +108,13 @@ def test_models_exposes_image_tts_and_music_separately() -> None:
             "id": "qwen3_tts",
             "name": "Qwen3 TTS",
             "provider": "wangp",
-            "capabilities": ["text_to_speech"],
+            "languages": ["English", "auto"],
+            "capabilities": [
+                "text_to_speech",
+                "voice_cloning",
+                "reference_transcript",
+                "language_selection",
+            ],
         }
     ]
     assert _models(session, "music") == [
@@ -176,6 +193,17 @@ def test_image_to_image_uses_reference_input_when_required(tmp_path: Path) -> No
     assert settings["image_refs"] == [str(image.resolve())]
     assert settings["resolution"] == "1024x768"
     assert settings["denoising_strength"] == 0.65
+    assert settings["image_mode"] == 1
+
+
+def test_text_to_image_selects_image_output_for_dual_output_models() -> None:
+    settings = _settings(
+        {
+            "operation": "text_to_image",
+            "params": {"model": "t2v_2_2", "prompt": "sunrise"},
+        }
+    )
+    assert settings["image_mode"] == 1
 
 
 def test_image_to_image_uses_model_control_mode(tmp_path: Path) -> None:
@@ -226,6 +254,23 @@ def test_music_settings_map_lyrics_style_and_duration() -> None:
     }
 
 
+def test_stable_audio_uses_description_as_its_main_prompt() -> None:
+    assert _settings(
+        {
+            "operation": "text_to_audio",
+            "params": {
+                "model": "stable_audio3_small",
+                "prompt": "soft rain and distant thunder",
+                "lyrics": "ignored lyrics",
+            },
+        },
+        {"base_model_type": "stable_audio3_small"},
+    ) == {
+        "model_type": "stable_audio3_small",
+        "prompt": "soft rain and distant thunder",
+    }
+
+
 def test_tts_settings_map_voice_clone_fields(tmp_path: Path) -> None:
     audio = tmp_path / "voice.wav"
     audio.write_bytes(b"wav")
@@ -246,17 +291,56 @@ def test_tts_settings_map_voice_clone_fields(tmp_path: Path) -> None:
         "prompt": "Hello",
         "alt_prompt": "Reference words",
         "model_mode": "English",
-        "speech_speed": 1.1,
         "audio_guide": str(audio.resolve()),
     }
 
 
+def test_tts_uses_voice_only_for_custom_voice_and_speed_only_for_index25() -> None:
+    custom_voice = _settings(
+        {
+            "operation": "tts_encoded",
+            "params": {
+                "model": "qwen3_tts_customvoice",
+                "text": "Hello",
+                "voice": "Ryan",
+                "language": "English",
+            },
+        },
+        {"base_model_type": "qwen3_tts_customvoice"},
+    )
+    assert custom_voice["model_mode"] == "Ryan"
+
+    index = _settings(
+        {
+            "operation": "tts_encoded",
+            "params": {
+                "model": "index_tts25",
+                "text": "Hello",
+                "voice": "unused",
+                "language": "EN",
+                "speed": 1.25,
+            },
+        },
+        {"base_model_type": "index_tts25"},
+    )
+    assert index["model_mode"] == "EN"
+    assert index["custom_settings"] == {"speech_speed": 1.25}
+
+
 def test_generated_path_requires_successful_existing_file(tmp_path: Path) -> None:
     video = tmp_path / "output.mp4"
     video.write_bytes(b"video")
     result = SimpleNamespace(success=True, generated_files=[str(video)], artifacts=[])
     assert _generated_path(result, "video") == str(video.resolve())
 
+    image = tmp_path / "wrong.png"
+    image.write_bytes(b"image")
+    with pytest.raises(RuntimeError, match="without a generated media file"):
+        _generated_path(
+            SimpleNamespace(success=True, generated_files=[str(image)], artifacts=[]),
+            "video",
+        )
+
     with pytest.raises(RuntimeError, match="denied"):
         _generated_path(
             SimpleNamespace(