ComfyUI-MiniMax-H3-API

ComfyUI MiniMax H3 API

ComfyUI nodes for the MiniMax H3 Video V2 API. The plugin supports MiniMax H3, MiniMax H3 Max, MiniMax H3 Max Turbo, and MiniMax H3 Fast generation, text-to-video, first/last-frame video, multimodal reference video, H3-Context-IR, 768P-to-2K regeneration, task management, download, and preview.

Configuration

Copy local.example.json to local.json and add your MiniMax API key:

{
  "api_key": "YOUR_MINIMAX_API_KEY",
  "base_url": "https://api.minimax.cn",
  "poll_interval": 5,
  "request_timeout": 60,
  "max_wait_seconds": 3600
}

local.json is ignored by Git. The API key is read when a node executes and is never included in workflow files or request outputs.

The supplied Max / Max Turbo API document (2026-09-06) specifies https://api.minimax.cn for China and https://api.minimax.io for overseas accounts. Use the matching regional key and endpoint. Existing explicit base_url settings remain in effect; update your local configuration when migrating.

Nodes

• MiniMax H3 Content Builder builds and validates the multimodal content array.
• MiniMax H3 Generate Video creates a video task and automatically polls until completion.
• MiniMax H3 Max Generate Video supports text-to-video, first-frame / first-and-last-frame generation, and multimodal image/video/audio references at 480P or 768P for 5–15 seconds.
• MiniMax H3 Max Turbo Generate Video is a separate node using MiniMax-H3-Max-Turbo. It supports text-to-video and first-frame / first-and-last-frame generation at 480P or 768P for 5–15 seconds. Reference images/videos/audio are not supported.
• MiniMax H3 Fast Generate Video focuses on generation quality at 480P for 4–15 seconds. It supports text-to-video, first/last-frame image-to-video, and multimodal reference generation; the required style1 subtype is applied automatically. Generation costs ¥0.30/second.
• MiniMax H3 Context IR creates an asynchronous prompt-enhancement task and returns reusable enhanced content.
• MiniMax H3 Regenerate 2K reuses an exact 768P generation request and adds the source video.
• MiniMax H3 Query Task queries a task created during the last seven days.
• MiniMax H3 List Tasks lists and filters recent tasks.
• MiniMax H3 Cancel Delete Task cancels queued tasks or deletes succeeded/failed task records.
• MiniMax H3 Preview Video downloads the result as video/MiniMax_日期_时间.mp4 and scales the preview with the node size.

Content Builder inputs

• Prompt only creates text-to-video content.
• Connecting first_frame, last_frame, or both creates frame-guided content and automatically uses the adaptive ratio.
• Connecting image_1–image_9, video_1–video_3, or audio_1–audio_3 creates reference content. At least one image or video is required.

The content mode is inferred from connected inputs. Native ComfyUI IMAGE, VIDEO, and AUDIO inputs are converted to MiniMax data URIs. The complete JSON request must not exceed 64 MB.

For Max and Max Turbo, a last frame requires a first frame. Frame inputs cannot be mixed with references. Max accepts at most 9 reference images, 3 reference videos, 3 reference audios, and 12 reference media items in total; audio alone is invalid. Each reference video/audio must be 2–15 seconds, with combined video duration at most 15 seconds and combined audio duration separately at most 15 seconds. Media duration/download/decode validation is performed by the API.

Both Max nodes expose optional prompt_expansion_mode (disabled, balanced, quality; default balanced) and seed. Seed 0 omits the API field for server-side random generation; 1–4294967295 sends a seed. Existing Max workflows can omit these new inputs. Neither model uses style, H3-Context-IR, or 2K regeneration.

Max pricing display

Rates below come from the supplied 2026-09-06 document. The node displays a standard-price estimate in CNY, or USD when base_url is https://api.minimax.io; actual account billing prevails. Max nodes do not convert charges to credits.

| Item | China CNY | Overseas USD | | --- | ---: | ---: | | Max 480P output / second | 0.33 | 0.05 | | Max 768P output / second | 0.50 | 0.08 | | Turbo 480P output / second | 0.17 | 0.025 | | Turbo 768P output / second | 0.26 | 0.04 | | Max reference video / input second, 480P output | 0.37 | 0.0553 | | Max reference video / input second, 768P output | 0.97 | 0.143 | | Max reference image beyond the first 2 / image | 0.50 | 0.074 |

Max cost = output seconds × output rate + input video seconds × video rate + max(image count − 2, 0) × image rate. Reference audio and prompt expansion incur no separate charge. total_seconds alone is insufficient to calculate Max cost; incomplete usage is reported instead of guessing. For example, Max 768P output of 5 seconds with 10 input video seconds, 3 images and 10 audio seconds costs CNY 12.70 or USD 1.904 at these rates.

MiniMax H3 Fast supports only 480P and does not support H3-Context-IR or 2K regeneration.

Polling is always enabled. AIGC watermarking is always disabled. These internal settings are not shown in the node UI.

Regeneration

Generate the source video at 768P, connect request_json from Generate Video to generation_request_json on Regenerate 2K, and provide the exact 768P source video. MiniMax requires the source video to contain audio and match its H3 768P output specification.

Examples

Import workflow JSON files from examples/. See examples/README.md for required inputs and usage notes. Example workflows never contain an API key.

API documentation

• Create video generation task
• Query task
• List tasks
• Cancel or delete task
• H3-Context-IR
• Video regeneration