MiniMax H3 Studio 👻:
LP-MiniMaxH3Studio combines material loading, prompt preparation, and MiniMax H3 video setup in one ComfyUI node. You can describe a video idea and provide reference materials through its integrated materials field; the node prepares an H3 prompt, then calls the appropriate built-in MiniMax H3 node to create the conditioning and latent needed for generation. It can use an optional language-model configuration to refine the prompt, but it can also pass your description through without that configuration. Depending on the selected mode, it supports image-guided frames or a collection of image, video, and audio references. The node prepares generation inputs; it does not render a finished video by itself.
MiniMax H3 Studio 👻 Input Parameters:
clip
Connect a CLIP model for the built-in MiniMax H3 node to process the prepared prompt. This required input contributes to the conditioning used later in the generation workflow. It is a model connection, so there is no numeric minimum, maximum, or default.
vae
Connect a VAE for the built-in MiniMax H3 node. The node passes it along when preparing the video latent and, where applicable, image or video references. This required input has no numeric range or default.
h3_materials
Provide the integrated materials data containing your prompt and any selected images, audio, videos, or reference-role descriptions. The node parses this STRING field and loads the referenced materials before building the H3 prompt. Its default is an empty string; no minimum or maximum length is specified. You need to include a usable video description, because the node stops if the resulting prompt is empty.
mode
Choose how your idea and materials guide the video. The node maps the selected option to an internal mode: text-to-video (T2VA), first-frame image guidance (I2VA), first- and last-frame guidance (FL2VA), last-frame guidance (L2VA), or reference-guided generation (Ref2VA). These modes affect both prompt preparation and which built-in H3 node receives your materials. The exact display labels and default for this COMBO input are defined in imported MODE_OPTIONS and are not present in the supplied source, so they cannot be listed reliably here.
style_preset
Choose a style for prompt preparation. The selected option is mapped to a style key used when constructing the language-model instructions; an unrecognized value maps to the general style. This setting does not directly change the latent-generation call. The exact available labels and default are defined in imported STYLE_OPTIONS and are not present in the supplied source.
duration
Set the requested video duration in seconds. The node converts it to a frame count and rounds that count upward to the frame grid expected by the built-in H3 node, so the resulting frame count may not correspond to precisely the entered duration. The default is 5.0 seconds, the minimum is 1.0, the maximum is 15.0, and the step is 0.1 second.
width
Set the requested video width passed to the built-in H3 node. Use it with height to choose the intended output dimensions. The default is 1344 pixels, the minimum is 32, the maximum is 16384, and the step is 32 pixels. The supplied code does not establish which dimensions your particular model or hardware can successfully generate.
height
Set the requested video height passed to the built-in H3 node. Together with width, it determines the requested frame dimensions. The default is 768 pixels, the minimum is 32, the maximum is 16384, and the step is 32 pixels. Support for a particular size depends on the connected generation setup and is not guaranteed by this node.
output_language
Select the language requested for language-model-generated prompt text. The source identifies English and 中文 as English and Chinese choices, respectively; the default is English. This setting affects the prompt-optimization request when an API key is available. Without an API key, the pipeline passes through your parsed prompt instead of translating or optimizing it. No complete option list beyond the supplied tooltip is available in the source.
audio_vae
Optionally connect an audio VAE for reference-guided generation. This connection becomes required when Ref2VA contains successfully loaded standalone audio or audio extracted from a reference video; otherwise, the node raises an error before calling the built-in reference node. No default model or numeric range is specified.
ref_image_size
Choose how reference-image size is handled in Ref2VA mode. The options are match, which adapts references to the generation resolution, and max, described in the node tooltip as using a 2048-pixel short side for best fidelity. The default is match. The setting is passed to the built-in reference node only in Ref2VA; it does not affect the other modes.
llm_config
Optionally connect an H3_LLM_CONFIG configuration to enable API-based prompt optimization. When the configuration supplies an API key, the pipeline uses its model, base URL, token limit, thinking setting, and temperature to request a prepared prompt. Without an API key, or if the API call fails, it passes through your parsed prompt and reference-role text instead. This input has no numeric range or default configuration specified in the node interface.
MiniMax H3 Studio 👻 Output Parameters:
positive
The CONDITIONING returned by the built-in MiniMax H3 node. It represents the prepared guidance for the subsequent generation workflow, based on the prompt and the selected mode. Connect it to the downstream component that expects positive conditioning; it is not a finished image or video.
latent
The LATENT returned by the built-in MiniMax H3 node. It is the prepared latent input for downstream generation at the requested dimensions and converted frame length. It is not a playable video, so further workflow steps are needed to generate and view the result.
prompt
The final STRING prompt passed to the built-in MiniMax H3 node. Inspect it to see what text guided the prepared conditioning: with a successful configured API call, it may contain structured, mode-specific sections; without an API key or after an API failure, it contains the parsed prompt and reference-role text passed through by the pipeline.
MiniMax H3 Studio 👻 Usage Tips:
- Start with a clear description in
h3_materials, even when adding references. The node cannot continue if prompt preparation produces an empty prompt. - Choose
I2VAto guide the first frame with the first loaded image,L2VAto guide the last frame with the first loaded image, orFL2VAto use the first two loaded images for the first and last frames. InFL2VA, supplying only one image changes the prompt’s frame-alignment instruction to first-frame guidance, while leaving the last-frame input empty. - Choose
Ref2VAwhen you want to pass loaded images, videos, or audio as references. Connectaudio_vaeif any successfully loaded reference includes audio, including audio from a video. - Use
ref_image_sizeonly when working inRef2VA. Begin with itsmatchdefault; considermaxwhen reference-image fidelity is more important to your workflow. - Connect
llm_configwhen you want API-based prompt optimization. If you leave it unconnected, make your own prompt sufficiently complete, because the node will pass it through rather than expand it with the language model. - Check the
promptoutput before a longer generation run. It lets you confirm the prepared wording, reference labels, and language that were actually passed to the built-in H3 node. - Set
duration,width, andheightfor the result you intend to generate, while accounting for the node’s upward frame-count alignment. A larger requested size or longer duration should be tested against the capabilities of your connected generation setup.
MiniMax H3 Studio 👻 Common Errors and Solutions:
提示词为空:请在提示词输入框中描述视频创意
- Explanation: The prepared H3 prompt is empty, so the node has no video description to send to the built-in H3 node.
- Solution: Enter a substantive video description in the prompt portion of
h3_materials, then confirm that the materials data reaches this node.
Ref2VA 模式包含音频素材,需要连接 audio_vae 输入
- Explanation:
Ref2VAfound loaded standalone audio or audio in a reference video, butaudio_vaeis not connected. - Solution: Connect a suitable audio VAE to
audio_vae, or remove the audio reference if you do not intend to use it.
未找到 ComfyUI 内置节点 MiniMaxH3ImageToVideo,请升级到包含 MiniMax H3 官方节点的 ComfyUI 版本
- Explanation: The selected non-
Ref2VAmode needs the built-in MiniMax H3 image-to-video node, but it is not registered in this ComfyUI installation. - Solution: Update to a ComfyUI version that includes the built-in MiniMax H3 nodes, then restart ComfyUI.
未找到 ComfyUI 内置节点 MiniMaxH3ReferenceToVideo,请升级到包含 MiniMax H3 官方节点的 ComfyUI 版本
- Explanation:
Ref2VAneeds the built-in MiniMax H3 reference-to-video node, but it is not registered in this ComfyUI installation. - Solution: Update to a ComfyUI version that includes the built-in MiniMax H3 reference node, then restart ComfyUI.
API 调用失败,返回原始 prompt
- Explanation: This is a warning from the prompt pipeline, not necessarily a stopped run. The API request did not return usable output, so the node falls back to your parsed prompt and reference-role text.
- Solution: Inspect the
promptoutput. If you expected optimization, check the API key and other settings supplied throughllm_config, then retry; keep a complete original prompt so the fallback remains useful.
Missing required input: clip, vae, h3_materials, mode, style_preset, duration, width, height, or output_language
- Explanation: ComfyUI may prevent execution if a required connection or field is unavailable. The exact validation wording depends on ComfyUI.
- Solution: Check each required socket and setting listed above. In particular, connect both
clipandvaeand ensure the materials field contains your intended description.
Built-in H3 node fails while processing references or generation settings
- Explanation: After preparation, LP-MiniMaxH3Studio calls a built-in H3 node. An error raised there may concern the connected models, loaded reference materials, requested dimensions, or other requirements of that built-in node; the supplied Studio source does not establish a single error message for these cases.
- Solution: Read the full ComfyUI error to identify the failing input. Confirm that referenced files load, that the connected models belong to your H3 workflow, and that your chosen dimensions and duration work with that setup.
