sgl-project/sglang · error · ValueError

ref2va requires at least one of reference_image, reference_v

Error message

ref2va requires at least one of reference_image, reference_video, or reference_audio

What it means

Raised by the ComfyUI SGLDiffusion MiniMax-H3 node when task='ref2va' (reference-to-video) but no reference_image, reference_video, or reference_audio is provided. It mirrors server-side validation locally so the user knows which input to wire.

Source

Thrown at python/sglang/multimodal_gen/apps/ComfyUI_SGLDiffusion/nodes.py:744

                    "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,

View on GitHub (pinned to 0132848349)

Solutions

1. Wire at least one of reference_image, reference_video, or reference_audio
2. If you have no references, use task 't2va' (text-only)
3. If you only have start/end frames, use task 'fl2va'

Example fix

// before
task='ref2va', no reference inputs connected
// after
task='ref2va', reference_image=load_image_output

required = ['reference_image','reference_video','reference_audio']
if task == 'ref2va' and not any(globals()[k] is not None for k in required):
    raise UserError('ref2va needs a reference input')

try: node.generate(...) except ValueError as e: show_user(str(e))

When it happens

Trigger: Calling generate() with task='ref2va' and all of reference_image, reference_video, reference_audio unset (empty conditions).

Common situations: Choosing ref2va in ComfyUI without attaching any reference nodes; assuming the prompt alone is enough after switching tasks.

Related errors

• fl2va requires first_frame, last_frame, or both
• t2va takes no conditioning inputs; pick another task
• Failed to generate MiniMax-H3 video: {str(e)}
• img_position_ids must be [1, S, 3], got {list(img_position_i
• MiniMax H3 model variant must be a non-empty string

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/17a4976cb0dc05c7. Report an issue: GitHub.