diff --git a/python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-add-model/SKILL.md b/python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-add-model/SKILL.md
index 4f0bb45c1..85450d4af 100644
--- a/python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-add-model/SKILL.md
+++ b/python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-add-model/SKILL.md
@@ -574,6 +574,39 @@ After implementation, **you must verify that the generated output is not noise**
 2. Running the Diffusers pipeline and SGLang pipeline side-by-side with the same seed
 3. Checking each stage's output shape and value range independently
 
+### Step 10: Decide the ComfyUI Route (Optional)
+
+A model is reachable from ComfyUI two ways. Pick one deliberately — the wrong
+choice costs several hundred lines of weight-mapping code that buys nothing.
+
+**Server route.** ComfyUI sends an HTTP request and SGLang runs the whole
+pipeline. Choose this when the model needs conditioning ComfyUI cannot supply
+(audio, reference materials, task routing), produces more than one modality,
+or has its own request contract.
+
+Cost: nothing, if the request fits the existing `generate_image` /
+`generate_video` fields. If the model has extra request fields, pass them
+through `extra_fields` — the request schemas accept unknown keys, so the
+client in `apps/ComfyUI_SGLDiffusion/core/server_api.py` does **not** need a
+per-model change. Add a node in `nodes.py` only when the inputs are worth
+surfacing as ComfyUI widgets. `SGLDiffusionGenerateH3` is the worked example.
+
+**Executor route.** ComfyUI's KSampler drives the denoise loop and SGLang
+replaces the DiT forward, using ComfyUI's own text encoders and VAE. Choose
+this only when the model denoises a single latent tensor that ComfyUI already
+knows how to build and decode.
+
+Cost, per model: a `runtime/pipelines/comfyui_<model>_pipeline.py` that maps
+ComfyUI's single-file checkpoint layout onto the native module tree (350-690
+lines in the existing three), an executor in
+`apps/ComfyUI_SGLDiffusion/executors/` that adapts latent layout and
+conditioning to `Req`, and entries in both dicts in `core/generator.py`.
+
+The deciding question is not model size or modality — it is whether ComfyUI's
+sampler can drive the model's loop unchanged. If reproducing the conditioning
+inside ComfyUI would duplicate stages the server already runs, take the server
+route.
+
 ## Reference Implementations
 
 ### Hybrid Style (recommended for most new models)
diff --git a/python/sglang/multimodal_gen/apps/ComfyUI_SGLDiffusion/README.md b/python/sglang/multimodal_gen/apps/ComfyUI_SGLDiffusion/README.md
index 305190b68..1fb059efe 100644
--- a/python/sglang/multimodal_gen/apps/ComfyUI_SGLDiffusion/README.md
+++ b/python/sglang/multimodal_gen/apps/ComfyUI_SGLDiffusion/README.md
@@ -16,6 +16,7 @@ The plugin supports two modes of operation: **Server Mode** (via HTTP API) and *
 - **Z-Image**: High-speed image generation models (e.g., `Z-Image-Turbo`)
 - **FLUX**: State-of-the-art text-to-image models (e.g., `FLUX.1-dev`)
 - **Qwen-Image**: Multi-modal image generation models (e.g., `Qwen-Image`,`Qwen-Image-2512`). *Note: Image editing support is currently experimental and may have some issues.*
+- **MiniMax-H3**: Joint video-and-audio generation, server mode only (`SGLDiffusion Generate H3`)
 
 ### Mode 1: Server Mode (HTTP API)
 Connect to a standalone SGLang Diffusion server.
@@ -35,6 +36,35 @@ Leverage SGLang's high-performance sampling directly within ComfyUI while using
 3. **Sample**: Connect the loaded model to standard ComfyUI samplers. SGLang will handle the sampling process efficiently.
 4. **LoRA Support**: Use the `SGLDiffusion LoRA Loader` for native LoRA integration.
 
+## Adding a Model
+
+Pick the mode before writing code; the wrong one costs several hundred lines
+of weight mapping that buys nothing.
+
+Take **Server Mode** when the model needs conditioning ComfyUI cannot supply
+(audio, reference materials, task routing), emits more than one modality, or
+has its own request contract. Reproducing that inside ComfyUI would duplicate
+stages the server already runs.
+
+- If the request fits the existing `generate_image` / `generate_video` fields,
+  there is nothing to write — point the existing nodes at the server.
+- If the model has extra request fields, pass them via `extra_fields`. The
+  request schemas accept unknown keys, so `core/server_api.py` needs no
+  per-model change.
+- Add a node in `nodes.py` only to surface those inputs as ComfyUI widgets.
+  `SGLDiffusionGenerateH3` is the worked example.
+
+Take **Integrated Mode** only when the model denoises a single latent tensor
+that ComfyUI already knows how to build and decode, so its KSampler can drive
+the loop unchanged. Each model then needs:
+
+- `runtime/pipelines/comfyui_<model>_pipeline.py` mapping ComfyUI's
+  single-file checkpoint layout onto the native module tree (350-690 lines in
+  the existing three)
+- an executor in `executors/` adapting latent layout and conditioning to `Req`
+- entries in both `pipeline_class_dict` and `executor_class_dict` in
+  `core/generator.py`
+
 ## Example Workflows
 
 Reference workflow files are provided in the `workflows/` directory:
diff --git a/python/sglang/multimodal_gen/apps/ComfyUI_SGLDiffusion/core/server_api.py b/python/sglang/multimodal_gen/apps/ComfyUI_SGLDiffusion/core/server_api.py
index 2f86efe79..0b3f92bec 100644
--- a/python/sglang/multimodal_gen/apps/ComfyUI_SGLDiffusion/core/server_api.py
+++ b/python/sglang/multimodal_gen/apps/ComfyUI_SGLDiffusion/core/server_api.py
@@ -219,6 +219,7 @@ class SGLDiffusionServerAPI:
         generator_device: Optional[str] = "cuda",
         input_reference: Optional[str] = None,
         output_path: Optional[str] = None,
+        extra_fields: Optional[Dict[str, Any]] = None,
     ) -> Dict[str, Any]:
         """
         Generate a video using SGLang Diffusion API and wait for completion.
@@ -238,6 +239,10 @@ class SGLDiffusionServerAPI:
             enable_teacache: Enable TEA cache acceleration
             generator_device: Device for random generator ("cuda" or "cpu")
             input_reference: Path to input reference image for image-to-video
+            extra_fields: Model-specific request fields merged into the payload
+                last, so a caller can reach a model's own request surface
+                (MiniMax-H3's `task`/`conditions`/`target`, per-model flow
+                shifts) without this client growing a parameter per model
 
         Returns:
             Dictionary containing completed video job information with file_path
@@ -281,6 +286,10 @@ class SGLDiffusionServerAPI:
             payload["input_reference"] = input_reference
         if output_path:
             payload["output_path"] = output_path
+        # merged last so a model-specific field wins over a generic default of
+        # the same name (H3 sizes its output from `target`, not `size`)
+        if extra_fields:
+            payload.update(extra_fields)
 
         try:
             # Create video generation job
diff --git a/python/sglang/multimodal_gen/apps/ComfyUI_SGLDiffusion/nodes.py b/python/sglang/multimodal_gen/apps/ComfyUI_SGLDiffusion/nodes.py
index af31fe058..f3d2275ee 100644
--- a/python/sglang/multimodal_gen/apps/ComfyUI_SGLDiffusion/nodes.py
+++ b/python/sglang/multimodal_gen/apps/ComfyUI_SGLDiffusion/nodes.py
@@ -576,6 +576,222 @@ class SGLDiffusionGenerateVideo:
         return (video, video_path)
 
 
+class SGLDiffusionGenerateH3:
+    """Node to generate joint video and audio with MiniMax-H3.
+
+    H3 denoises a packed video+audio sequence in one pass and routes its
+    conditioning by task rather than by a single reference slot, so it needs
+    its own request shape (`task` / `conditions` / `target`) that the generic
+    video node does not model. The returned MP4 carries both streams.
+    """
+
+    TASKS = ["t2va", "fl2va", "ref2va"]
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "sgld_client": ("SGLD_CLIENT",),
+                "positive_prompt": (
+                    "STRING",
+                    {
+                        "default": "",
+                        "tooltip": "Text prompt. Reference material is addressed "
+                        "positionally as <Picture 1>, <Video 1>, <Audio 1>",
+                    },
+                ),
+                "task": (
+                    cls.TASKS,
+                    {
+                        "default": "t2va",
+                        "tooltip": "t2va: text only. fl2va: first/last keyframes. "
+                        "ref2va: image, video, and audio references",
+                    },
+                ),
+            },
+            "optional": {
+                "first_frame": (
+                    "IMAGE",
+                    {"tooltip": "fl2va: becomes the clip's first frame"},
+                ),
+                "last_frame": (
+                    "IMAGE",
+                    {"tooltip": "fl2va: becomes the clip's last frame"},
+                ),
+                "reference_image": (
+                    "IMAGE",
+                    {
+                        "tooltip": "ref2va: guides identity and style; not "
+                        "preserved as an endpoint frame"
+                    },
+                ),
+                "reference_video": (
+                    "STRING",
+                    {"default": "", "tooltip": "ref2va: path or URL to a video"},
+                ),
+                "reference_audio": (
+                    "STRING",
+                    {"default": "", "tooltip": "ref2va: path or URL to audio"},
+                ),
+                "negative_prompt": ("STRING", {"default": ""}),
+                "seed": ("INT", {"default": 1101, "min": -1, "max": 2**32 - 1}),
+                "steps": ("INT", {"default": 50, "min": 1, "max": 100}),
+                "short_edge": ("INT", {"default": 768, "min": 256, "max": 1536}),
+                "aspect_ratio": (
+                    ["16:9", "9:16", "1:1", "auto"],
+                    {"default": "16:9"},
+                ),
+                "duration_seconds": (
+                    "FLOAT",
+                    {"default": 5.0, "min": 4.0, "max": 15.0, "step": 0.5},
+                ),
+                "flow_shift": ("FLOAT", {"default": 12.0, "min": 0.0, "max": 30.0}),
+                "audio_flow_shift": (
+                    "FLOAT",
+                    {"default": 3.0, "min": 0.0, "max": 30.0},
+                ),
+            },
+        }
+
+    RETURN_TYPES = ("VIDEO", "STRING")
+    RETURN_NAMES = ("video", "video_path")
+    FUNCTION = "generate"
+    CATEGORY = "SGLDiffusion"
+    OUTPUT_NODE = False
+
+    @staticmethod
+    def _material_uri(value: str) -> str:
+        """Local paths become file:// URIs; remote URLs are passed through."""
+        if value.startswith(("http://", "https://", "file://")):
+            return value
+        return f"file://{os.path.abspath(value)}"
+
+    def generate(
+        self,
+        sgld_client: SGLDiffusionServerAPI,
+        positive_prompt: str,
+        task: str,
+        first_frame: torch.Tensor = None,
+        last_frame: torch.Tensor = None,
+        reference_image: torch.Tensor = None,
+        reference_video: str = "",
+        reference_audio: str = "",
+        negative_prompt: str = "",
+        seed: int = 1101,
+        steps: int = 50,
+        short_edge: int = 768,
+        aspect_ratio: str = "16:9",
+        duration_seconds: float = 5.0,
+        flow_shift: float = 12.0,
+        audio_flow_shift: float = 3.0,
+    ):
+        """Build H3's task-shaped request and submit it through the server API."""
+        if not positive_prompt:
+            raise ValueError("Prompt cannot be empty")
+
+        # 1. keyframes carry a frame_index and are preserved as endpoints;
+        #    references are semantic material and keep request order, because
+        #    the prompt addresses them positionally per modality
+        conditions = []
+        if first_frame is not None:
+            conditions.append(
+                {
+                    "type": "image",
+                    "uri": self._material_uri(get_image_path(first_frame)),
+                    "role": "keyframe",
+                    "frame_index": 0,
+                }
+            )
+        if last_frame is not None:
+            conditions.append(
+                {
+                    "type": "image",
+                    "uri": self._material_uri(get_image_path(last_frame)),
+                    "role": "keyframe",
+                    "frame_index": -1,
+                }
+            )
+        if reference_image is not None:
+            conditions.append(
+                {
+                    "type": "image",
+                    "uri": self._material_uri(get_image_path(reference_image)),
+                    "role": "reference",
+                }
+            )
+        if reference_video:
+            conditions.append(
+                {
+                    "type": "video",
+                    "uri": self._material_uri(reference_video),
+                    "role": "reference",
+                }
+            )
+        if reference_audio:
+            conditions.append(
+                {
+                    "type": "audio",
+                    "uri": self._material_uri(reference_audio),
+                    "role": "reference",
+                }
+            )
+
+        # 2. reject wiring the server would reject anyway, but name the input
+        #    the user has to change
+        if task == "fl2va" and not (first_frame is not None or last_frame is not None):
+            raise ValueError("fl2va requires first_frame, last_frame, or both")
+        if task == "ref2va" and not conditions:
+            raise ValueError(
+                "ref2va requires at least one of reference_image, "
+                "reference_video, or reference_audio"
+            )
+        if task == "t2va" and conditions:
+            raise ValueError("t2va takes no conditioning inputs; pick another task")
+
+        # 3. `target` resolves the aligned canvas and frame count; the `size`
+        #    the server API always sends is unused by H3
+        extra_fields = {
+            "task": task,
+            "conditions": conditions,
+            "target": {
+                "short_edge": short_edge,
+                "aspect_ratio": aspect_ratio,
+                "duration_seconds": duration_seconds,
+            },
+            "flow_shift": flow_shift,
+            "audio_flow_shift": audio_flow_shift,
+        }
+
+        request_params = {
+            "prompt": positive_prompt,
+            "seconds": int(duration_seconds),
+            "num_inference_steps": steps,
+            "output_path": folder_paths.get_temp_directory(),
+            "extra_fields": extra_fields,
+        }
+        if negative_prompt:
+            request_params["negative_prompt"] = negative_prompt
+        if seed >= 0:
+            request_params["seed"] = seed
+
+        try:
+            response = sgld_client.generate_video(**request_params)
+        except Exception as e:
+            raise RuntimeError(f"Failed to generate MiniMax-H3 video: {str(e)}")
+
+        video_path = response.get("file_path", "")
+        # H3 aligns the canvas server-side, so the resolved size is only known
+        # from the response; short_edge and aspect_ratio cannot reconstruct it
+        resolved_size = response.get("size", "")
+        if resolved_size:
+            width, height = (int(v) for v in resolved_size.split("x"))
+        else:
+            width = height = short_edge
+        video = convert_video_to_comfy_video(video_path, height, width)
+
+        return (video, video_path)
+
+
 class SGLDiffusionServerSetLora:
     """Node to set LoRA adapter for SGLang Diffusion server."""
 
@@ -696,6 +912,7 @@ NODE_CLASS_MAPPINGS = {
     "SGLDiffusionServerModel": SGLDiffusionServerModel,
     "SGLDiffusionGenerateImage": SGLDiffusionGenerateImage,
     "SGLDiffusionGenerateVideo": SGLDiffusionGenerateVideo,
+    "SGLDiffusionGenerateH3": SGLDiffusionGenerateH3,
     "SGLDiffusionServerSetLora": SGLDiffusionServerSetLora,
     "SGLDiffusionServerUnsetLora": SGLDiffusionServerUnsetLora,
     "SGLDUNETLoader": SGLDUNETLoader,
@@ -707,6 +924,7 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "SGLDiffusionServerModel": "SGLDiffusion Server Model",
     "SGLDiffusionGenerateImage": "SGLDiffusion Generate Image",
     "SGLDiffusionGenerateVideo": "SGLDiffusion Generate Video",
+    "SGLDiffusionGenerateH3": "SGLDiffusion Generate MiniMax-H3",
     "SGLDiffusionServerSetLora": "SGLDiffusion Server Set LoRA",
     "SGLDiffusionServerUnsetLora": "SGLDiffusion Server Unset LoRA",
     "SGLDUNETLoader": "SGLDiffusion UNET Loader",
diff --git a/python/sglang/multimodal_gen/apps/ComfyUI_SGLDiffusion/test/test_h3_request.py b/python/sglang/multimodal_gen/apps/ComfyUI_SGLDiffusion/test/test_h3_request.py
new file mode 100644
index 000000000..b46cf3b5e
--- /dev/null
+++ b/python/sglang/multimodal_gen/apps/ComfyUI_SGLDiffusion/test/test_h3_request.py
@@ -0,0 +1,290 @@
+"""Tests for the MiniMax-H3 ComfyUI node's request shape.
+
+These exercise the real node and the real server-API client together and mock
+only HTTP, so a change on either side of the payload contract fails here.
+"""
+
+import importlib.util
+import sys
+import types
+from pathlib import Path
+from unittest import mock
+
+import pytest
+import torch
+
+PLUGIN_DIR = Path(__file__).resolve().parents[1]
+PKG = "sgld_comfy_under_test"
+
+
+def _install_comfy_stubs() -> None:
+    """Stub the ComfyUI runtime modules the plugin imports at module scope.
+
+    The plugin only ever runs inside ComfyUI, so these packages are absent in
+    a plain checkout; stubbing them keeps the request-shape contract testable
+    without a ComfyUI install or a GPU.
+    """
+    folder_paths = types.ModuleType("folder_paths")
+    folder_paths.get_temp_directory = lambda: "/tmp"
+    sys.modules.setdefault("folder_paths", folder_paths)
+
+    comfy_api = types.ModuleType("comfy_api")
+    comfy_api_input = types.ModuleType("comfy_api.input")
+
+    class VideoInput:
+        pass
+
+    comfy_api_input.VideoInput = VideoInput
+    comfy_api.input = comfy_api_input
+    sys.modules.setdefault("comfy_api", comfy_api)
+    sys.modules.setdefault("comfy_api.input", comfy_api_input)
+
+    comfy = types.ModuleType("comfy")
+    comfy.model_detection = types.ModuleType("comfy.model_detection")
+    comfy.model_management = types.ModuleType("comfy.model_management")
+
+    comfy_utils = types.ModuleType("comfy.utils")
+    for name in (
+        "calculate_parameters",
+        "load_torch_file",
+        "state_dict_prefix_replace",
+        "unet_to_diffusers",
+    ):
+        setattr(comfy_utils, name, lambda *a, **k: None)
+    comfy.utils = comfy_utils
+
+    comfy_model_patcher = types.ModuleType("comfy.model_patcher")
+
+    class ModelPatcher:
+        def __init__(self, *a, **k):
+            pass
+
+    comfy_model_patcher.ModelPatcher = ModelPatcher
+    comfy.model_patcher = comfy_model_patcher
+
+    sys.modules.setdefault("comfy", comfy)
+    sys.modules.setdefault("comfy.model_detection", comfy.model_detection)
+    sys.modules.setdefault("comfy.model_management", comfy.model_management)
+    sys.modules.setdefault("comfy.utils", comfy_utils)
+    sys.modules.setdefault("comfy.model_patcher", comfy_model_patcher)
+
+
+def _load(module_name: str, relative_path: str):
+    """Load one plugin source file into the synthetic package."""
+    spec = importlib.util.spec_from_file_location(
+        module_name, PLUGIN_DIR / relative_path
+    )
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)
+    return module
+
+
+def _load_plugin():
+    """Load the node and its client without importing the sglang package root.
+
+    `sglang/__init__` pulls in the whole LLM serving stack, none of which this
+    contract depends on. `core.generator` is replaced by a placeholder because
+    it reaches for ComfyUI's model machinery and the node never calls it on the
+    server path.
+    """
+    _install_comfy_stubs()
+
+    package = types.ModuleType(PKG)
+    package.__path__ = [str(PLUGIN_DIR)]
+    sys.modules[PKG] = package
+
+    server_api = _load(f"{PKG}.core.server_api", "core/server_api.py")
+
+    core = types.ModuleType(f"{PKG}.core")
+    core.__path__ = [str(PLUGIN_DIR / "core")]
+    core.SGLDiffusionServerAPI = server_api.SGLDiffusionServerAPI
+    core.SGLDiffusionGenerator = object
+    sys.modules[f"{PKG}.core"] = core
+
+    _load(f"{PKG}.utils", "utils.py")
+    nodes = _load(f"{PKG}.nodes", "nodes.py")
+    return server_api, nodes
+
+
+SERVER_API, NODES = _load_plugin()
+SGLDiffusionServerAPI = SERVER_API.SGLDiffusionServerAPI
+SGLDiffusionGenerateH3 = NODES.SGLDiffusionGenerateH3
+
+RESOLVED_SIZE = "1344x768"
+
+
+class _Response:
+    def __init__(self, payload):
+        self._payload = payload
+
+    def raise_for_status(self):
+        pass
+
+    def json(self):
+        return self._payload
+
+
+def _run_node(**node_kwargs):
+    """Drive the node through the real client and capture the POST payload."""
+    client = SGLDiffusionServerAPI(base_url="http://127.0.0.1:30010")
+    captured = {}
+
+    def fake_post(url, json=None, headers=None, timeout=None):
+        captured.update(json)
+        return _Response({"id": "job-1"})
+
+    def fake_get(url, headers=None, timeout=None):
+        return _Response(
+            {
+                "id": "job-1",
+                "status": "completed",
+                "size": RESOLVED_SIZE,
+                "file_path": "/tmp/out.mp4",
+            }
+        )
+
+    node = SGLDiffusionGenerateH3()
+    with mock.patch(
+        f"{PKG}.core.server_api.requests.post",
+        side_effect=fake_post,
+    ), mock.patch(
+        f"{PKG}.core.server_api.requests.get",
+        side_effect=fake_get,
+    ), mock.patch(
+        f"{PKG}.nodes.get_image_path",
+        side_effect=lambda image: "/tmp/frame.png",
+    ):
+        result = node.generate(sgld_client=client, **node_kwargs)
+    return captured, result
+
+
+def _image():
+    return torch.zeros(1, 8, 8, 3)
+
+
+def test_t2va_sends_task_target_and_flow_shifts():
+    payload, _ = _run_node(positive_prompt="a cat", task="t2va")
+
+    assert payload["task"] == "t2va"
+    assert payload["conditions"] == []
+    assert payload["target"] == {
+        "short_edge": 768,
+        "aspect_ratio": "16:9",
+        "duration_seconds": 5.0,
+    }
+    assert payload["flow_shift"] == 12.0
+    assert payload["audio_flow_shift"] == 3.0
+
+
+def test_fl2va_maps_keyframes_to_frame_indices():
+    payload, _ = _run_node(
+        positive_prompt="continue the shot",
+        task="fl2va",
+        first_frame=_image(),
+        last_frame=_image(),
+    )
+
+    assert [c["role"] for c in payload["conditions"]] == ["keyframe", "keyframe"]
+    assert [c["frame_index"] for c in payload["conditions"]] == [0, -1]
+    assert all(c["uri"].startswith("file:///") for c in payload["conditions"])
+
+
+def test_ref2va_preserves_modality_order_for_prompt_tags():
+    payload, _ = _run_node(
+        positive_prompt="use <Picture 1> and <Audio 1>",
+        task="ref2va",
+        reference_image=_image(),
+        reference_video="/data/clip.mp4",
+        reference_audio="/data/voice.mp3",
+    )
+
+    assert [c["type"] for c in payload["conditions"]] == ["image", "video", "audio"]
+    assert {c["role"] for c in payload["conditions"]} == {"reference"}
+
+
+def test_remote_reference_urls_pass_through_unchanged():
+    url = "https://example.com/clip.mp4"
+    payload, _ = _run_node(
+        positive_prompt="follow <Video 1>",
+        task="ref2va",
+        reference_video=url,
+    )
+
+    assert payload["conditions"][0]["uri"] == url
+
+
+def test_node_reports_the_server_resolved_canvas():
+    _, (video, video_path) = _run_node(positive_prompt="a cat", task="t2va")
+
+    assert video.get_dimensions() == (1344, 768)
+    assert video_path == "/tmp/out.mp4"
+
+
+@pytest.mark.parametrize(
+    "kwargs,message",
+    [
+        ({"task": "fl2va"}, "fl2va requires"),
+        ({"task": "ref2va"}, "ref2va requires"),
+        ({"task": "t2va", "reference_video": "/data/clip.mp4"}, "t2va takes no"),
+    ],
+)
+def test_task_and_conditioning_must_agree(kwargs, message):
+    with pytest.raises(ValueError, match=message):
+        _run_node(positive_prompt="a cat", **kwargs)
+
+
+def test_extra_fields_win_over_generic_defaults():
+    """A model's own field must not be shadowed by a same-named generic default."""
+    client = SGLDiffusionServerAPI(base_url="http://127.0.0.1:30010")
+    captured = {}
+
+    def fake_post(url, json=None, headers=None, timeout=None):
+        captured.update(json)
+        return _Response({"id": "job-1"})
+
+    with mock.patch(
+        f"{PKG}.core.server_api.requests.post",
+        side_effect=fake_post,
+    ), mock.patch(
+        f"{PKG}.core.server_api.requests.get",
+        side_effect=lambda *a, **k: _Response(
+            {"id": "job-1", "status": "completed", "size": RESOLVED_SIZE}
+        ),
+    ):
+        client.generate_video(
+            prompt="a cat",
+            size="720x1280",
+            extra_fields={"size": "1344x768", "task": "t2va"},
+        )
+
+    assert captured["size"] == "1344x768"
+    assert captured["task"] == "t2va"
+
+
+@pytest.mark.parametrize("task", ["t2va", "fl2va", "ref2va"])
+def test_payload_validates_against_the_server_request_model(task):
+    """The node's payload must satisfy the schema the server actually parses.
+
+    The other tests mock HTTP, so they would still pass if a field were
+    misnamed or mistyped. This one feeds the captured payload to
+    VideoGenerationsRequest, closing that gap without a running server.
+    """
+    from sglang.multimodal_gen.runtime.entrypoints.openai.protocol import (
+        VideoGenerationsRequest,
+    )
+
+    conditioning = {
+        "t2va": {},
+        "fl2va": {"first_frame": _image()},
+        "ref2va": {"reference_image": _image()},
+    }[task]
+    payload, _ = _run_node(positive_prompt="a cat", task=task, **conditioning)
+
+    request = VideoGenerationsRequest(**payload)
+
+    # the H3 fields ride through as extras; losing them silently would leave a
+    # valid request that generates the wrong thing
+    assert request.task == task
+    assert request.target["short_edge"] == payload["target"]["short_edge"]
+    assert len(request.conditions) == len(payload["conditions"])