ComfyUI Node
MiniMax H3 Context IR (OpenAI) - leoleexh
Analyzes connected text/image/video/audio context and renders a validated MiniMax H3 Base or Ref2VA prompt. AI connection settings are editable on the node. Warning: a node-entered API key may be saved in workflow JSON. Connect the same media to the native H3 generation node.
MiniMax H3 Context IR (OpenAI) - leoleexh
- creative_profile
- first_frame
- last_frame
- ref_images
- ref_videos
- ref_video_audios
- ref_audios
- h3_prompt
- context_ir_json
- analysis_report
- warnings
- usage_json
◄user_prompt—►
◄modeauto►
◄width—►
◄height—►
◄length—►
◄model_presetgpt-5.6-sol►
◄custom_model►
◄api_key►
◄base_urlhttps://api.openai.com/v1►
◄api_modeauto►
◄reasoning_effortmedium►
◄output_languageen►
◄strictnessbalanced►
◄generate_background_musictrue►
◄generate_dialoguetrue►
◄generate_subtitlesfalse►
◄privacy_modecloud►
◄cache_enabledtrue►
◄max_visual_inputs40►
◄max_video_frames32►
◄transcription_modelgpt-4o-transcribe►
◄timeout_seconds120►
◄max_retries2►
◄generate_narrationfalse►
◄dialogue_languageauto►
Categoryleoleexh/MiniMax H3
Inputs (32)
| Name | Type | Default | Description |
|---|---|---|---|
| user_prompt | STRING | Original intent in natural language. Common forms such as `she says: Hello. English dialogue` are converted automatically; explicit <d>[Language] text</d> is also supported. | |
| mode | COMBO | auto | 6 options: auto, t2va, i2va, fl2va, l2va, ref2va |
| width | INT | Connect the same width INT used by the native H3 node, normally from the official Resolution Selector. | |
| height | INT | Connect the same height INT used by the native H3 node, normally from the official Resolution Selector. | |
| length | INT | Connect the exact same H3-aligned frame-count INT used by the native H3 node. In the official workflow this is the integer output of the duration Math Expression. | |
| model_preset | COMBO | gpt-5.6-sol | 5 options: gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, custom |
| custom_model | STRING | Used when model_preset is custom. Any compatible provider model ID is accepted. | |
| api_key | STRING | Type the key here, or leave blank to use an environment variable. ComfyUI's standard string widget shows this value and may save it in workflow JSON. Never share a workflow or screenshot containing a real key. | |
| base_url | STRING | https://api.openai.com/v1 | OpenAI-compatible API base URL. This is stored in workflow JSON; never put a key here. |
| api_mode | COMBO | auto | Auto tries Responses first and falls back only when the endpoint does not implement it. |
| reasoning_effort | COMBO | medium | 3 options: low, medium, high |
| output_language | COMBO | en | 3 options: en, zh, auto |
| strictness | COMBO | balanced | 3 options: conservative, balanced, creative |
| generate_background_music | BOOLEAN | true | ON asks for a fitting non-diegetic score unless the user explicitly requests no music. OFF forces the final prompt to prohibit background music. |
| generate_dialogue | BOOLEAN | true | ON allows natural dialogue when the scene supports it and preserves extracted or explicit lines. OFF forces no character dialogue or vocal performance. Narration is controlled separately. |
| generate_subtitles | BOOLEAN | false | ON asks H3 to burn in synchronized subtitles for actual dialogue. OFF prohibits dialogue captions. Generated-video text may still be imperfect. |
| privacy_mode | COMBO | cloud | Cloud sends selected media derivatives/audio/text to the endpoint. local_only uploads nothing. |
| cache_enabled | BOOLEAN | true | — |
| max_visual_inputs | INT | 401–64 | Global cloud-analysis budget: still images count as 1 each; the remaining slots are used for timestamped sampled frames from the single reference video. It must be at least the number of connected visual assets. |
| max_video_frames | INT | 322–32 | Hard cap for AI video frames. Actual sampling is adaptive at about 2 fps and is also limited by max_visual_inputs; the full frame batch is never uploaded. |
| transcription_model | STRING | gpt-4o-transcribe | — |
| timeout_seconds | INT | 12010–900 | — |
| max_retries | INT | 20–6 | — |
| generate_narration | BOOLEAN | false | OFF prohibits narration, voice-over, spoken descriptions, and reading prompt text aloud. ON only permits narration when the user_prompt explicitly requests it. |
| dialogue_language | COMBO | auto | Auto only permits context-appropriate dialogue. Selecting a language requires cloud planning to author at least one line in that language. Natural input such as `她说：你好。中文对白` is converted automatically; explicit <d>[Language] ...</d> also remains supported. |
| creative_profileopt | M3_CREATIVE_PROFILE | Optional connection from MiniMax H3 Creative Director. The structured profile is expanded against this node's exact duration, aspect ratio, mode, and connected references. | |
| first_frameopt | IMAGE | — | |
| last_frameopt | IMAGE | — | |
| ref_imagesopt | COMFY_AUTOGROW_V3 | — | |
| ref_videosopt | COMFY_AUTOGROW_V3 | — | |
| ref_video_audiosopt | COMFY_AUTOGROW_V3 | — | |
| ref_audiosopt | COMFY_AUTOGROW_V3 | — |
Outputs (5)
| Name | Type | Description |
|---|---|---|
| h3_prompt | STRING | — |
| context_ir_json | STRING | — |
| analysis_report | STRING | — |
| warnings | STRING | — |
| usage_json | STRING | — |