SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsminimax-h3-video

minimax-h3-video

Agents & models

View repository

Updated 10-16 08:41

About this skill

MiniMax H3 Cloud Video

Invoke the verified local orchestrator through the platform entry point in the plugin root directory. On Windows use scripts/h3-cloud.cmd、scripts/h3-onboard.cmd and scripts/h3-provision.cmd; on macOS/Linux use the script with the same name without an extension. The positions relative to this file are all ../../scripts/; before execution, select the entry point according to the current operating system and resolve it to an absolute path. The scripts/h3-* below all refer to the already-resolved corresponding platform entry point; do not call /bin/sh on Windows or run it via WSL.

Operating Principles

  • By default, officially use the Youyun ComfyUI Pure Edition Runtime Profile. Community images serve only as regression evidence; do not write image IDs, ComfyUI addresses, or workflow nodes into the Skill. At runtime, read the Deployment Binding and Runtime Spec.
  • Do not set up a separate welcome page or single-option platform page. Only briefly introduce the five modes when the user actively asks about the plugin's capabilities; for initial configuration, directly state that "Youyun Smart Computing" is currently in use, and put platform registration, API Key configuration, and credential entry in the same step.
  • The entire first-time flow advances one step at a time. Each turn presents only the one step that must currently be completed and one clear action; when the current step is not complete, do not preview or unfold subsequent steps. The order is fixed as: API Key configuration → credential verification and GPU selection → prompts/materials and parameters → single payment confirmation → generation and results. The system may internally execute checks that do not require user decisions consecutively, but must not dump the entire flow on the user all at once.
  • First-time use must be progressively disclosed. If the local credentials have not yet been configured or verified, the current turn may only complete the API Key guidance and end after asking the user to reply "completed". At this point, do not introduce GPU, video mode, prompt format, aspect ratio, duration, resolution, cost, or generation steps, and do not ask the user to provide a prompt. Only after credential verification succeeds in the next turn may GPU selection be entered.
  • Youyun credentials are managed by the official compshare-cli local profile. Do not read, print, copy, or ask the user to send the API Key.
  • This plugin does not distribute model weights. The community license for MiniMax H3 currently excludes regions such as the European Union, the United Kingdom, South Korea, and the United States; do not infer the user's location based on IP, time zone, or language. If the user is explicitly located in an excluded region or is uncertain about authorization, do not start paid resources; ask them to first obtain separate authorization from MiniMax and read the full license:https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE。
  • When the instance has not yet been established, first run the read-only h3-provision plan; only after the user confirms the selected region, inventory, and hourly price returned by the API, run h3-provision create --yes。
  • The user explicitly requests video generation, authorizing one corresponding paid generation task; before starting, confirm only the video mode, aspect ratio, duration, resolution, and estimated cost of this generation. The estimated cost must be based on the price in the latest read-only plan and the estimated duration of similar tasks, and must be clearly stated as an estimate. When generation is not explicitly requested, run only the read-only status or plan。
  • When the user states that an image must become the opening frame of the video, use first-frame keyframe mode and pass the absolute path to --first-frame-image。
  • When the user states that an image must become the ending frame of the video, use last-frame keyframe mode and pass the absolute path to --last-frame-image。
  • When the user specifies both the start and end frames, pass both --first-frame-image and --last-frame-image. The three keyframe modes reuse the same FL2VA Turbo 8-step workflow, with no additional model downloads.
  • When the user states that an image is only used to preserve a character, subject, or visual style, and does not require it to become the first or last frame, use the independent reference image mode; pass each image once --reference-image, supporting 1–9 images in parameter order. When the user only says "reference image" and the intent is unclear, first confirm whether it is a keyframe or a character/subject/style reference; reference image mode cannot be used together with any keyframe parameters.
  • First frame and first-and-last frame have already passed Youyun real-machine semantic verification; single last frame is still a candidate capability and requires one low-cost real-machine acceptance test before official release.
  • Reference images support 1–9 images and use the official Ref2VA native 20-step; both single-image and multi-image have completed Youyun real-machine semantic verification.
  • By default, have the orchestrator set fallback shutdown, and shut down immediately after success, failure, or interruption. Unless the user explicitly requests to keep the instance, do not pass --keep-running。
  • New users must actively choose among the three parallel configurations, with no preselection, no use of a "recommended option" label, and no silent fallback between configurations. Each item shows only the hardware, price, one factual basis, and one necessary limitation. Users may also request to view all current single-card specifications on the platform.
  • The current three configurations are 5090-96gb、5090-64gb、4090-64gb. The first two continue as the official test baseline;4090-64gb is an unverified low-price experimental item, and before payment it must be clearly stated that there is no guarantee of H3 speed and stability. Do not download or copy the 42.5GB model on site.
  • generate will automatically execute an idempotent Runtime Prepare before submission: verify the size and SHA-256 of the current six model/LoRA files, and only create safe symlinks; subsequent quick rechecks. Do not ask users to log in via SSH, manually operate ComfyUI, or re-download models already available on the platform.
  • Once ComfyUI returns prompt_id, the manifest will be persisted immediately. If an SSH/HTTP disconnection occurs afterward, the orchestrator only rebuilds the channel and resumes querying and downloading the original prompt_id , and never automatically submits a second paid task.

Video parameters

When the user is preparing a prompt, first remind them in one sentence to try to specify at the same time:

  • the visual content and motion;
  • the aspect ratio, for example 16:9 横屏、9:16 竖屏 or 1:1；
  • the duration, currently supporting 2–15 seconds;
  • the resolution, currently supporting 0.2–1.0MP.

An example can be given directly:一只橘猫坐在海边看日落，微风吹动毛发，电影感；16:9 横屏；10 秒；0.98MP（约 768P）。

If the user provides a first-frame image, the prompt should focus on describing "the motion, camera, and sound that occur after this image begins," and does not need to repeat the static composition already determined in the image. PNG, JPG, JPEG, and WebP are supported; local files are verified before GPU startup and uploaded to ComfyUI's independent input directory after the task starts. This mode reuses the same H3 model and Turbo workflow, with no additional model downloads.

If the user provides a last-frame image, the prompt should focus on describing "how the image naturally moves to this ending frame." If both a first frame and a last frame are provided, the prompt should focus on describing the action, camera, and scene transition between the two images. Both the first and last images are verified before GPU startup and uploaded separately to the task's independent directory.

If the user provides a character/subject/style reference image, the prompt describes the features to be preserved as well as the action, camera, and sound in the new scene, without requiring the composition to start from the original image. The orchestrator will automatically add, in parameter order, the same number of <Picture 1>…<Picture N> tag; do not write tags manually, and do not change the reference image into a first-frame input. This mode uses an independent Ref2VA main model, but maps it from the platform cache with zero copying, so users are not required to download it on the spot.

If the user has already explicitly written the aspect ratio, duration, and resolution, parse them directly into structured parameters and generate, without asking the user to choose a preset again. When only one of these is missing, ask only about the missing item; descriptions such as "high definition" or "as clear as possible" that cannot uniquely determine a numeric value still require the user to choose.

MP is the total pixel count, and P is the frame height; both change with the aspect ratio. The following is only for 16:9 intuitive reference for landscape:

Resolution Current workflow outputs approximately Common understanding
0.4MP 864×480 About 480P
0.6MP 1056×608 About 600P
0.8MP 1216×672 About 670P
0.9MP 1280×736 Close to 720P
0.98MP 1344×768 About 768P
1.0MP 1376×768 About 768P, slightly wider horizontally

When the user explicitly writes 480P、720P or 768P , it can be converted according to the table above in combination with the aspect ratio; before conversion, state the corresponding MP in one sentence. The old preview、balanced、short_hq is retained only as an internal compatibility shortcut and is not a required step for users. The 64GB Runtime does not run 15-second 0.9MP or higher configurations.

Action

  1. During initial configuration, directly state that UCloud is currently being used, and provide only two clickable entries:Register or log in、Create or manage API Key. At the same time, use one sentence to note that MiniMax H3 is subject to theCommunity Licenseand regional restrictions; do not separately display a welcome page, platform selection page, or text such as "more platforms will be supported in the future."
  2. If there is no isolated credential configuration on this machine yet, have the user run compshare config set --name <隔离名称> --no-activatein the terminal themselves, and enter the API Key through the official hidden input. The plugin must not read or echo the key. Remind the user not to send the API Key into the conversation, and ask them to reply only "Done" after completing it; then immediately end the current turn. This end point is a hard boundary, and you must not continue to display the GPU configuration page, generation mode, parameter examples, cost explanations, or ask for prompts. In the next turn, verify the credential first; on success, report only briefly "UCloud is connected," then proceed to step 3.
  3. After the credential succeeds, run scripts/h3-onboard options --platform compshare --credential-profile <隔离名称>. Prefer to use request_user_input to show the user a single-choice GPU question, with the three returned items as the three options in their original order; merge the option descriptions with priceDisplay、reason and the necessary caveat, with no preselection. If the client automatically provides "Other," clearly tell the user that it can be used to raise configuration requirements or view more; only when that tool is unavailable should it fall back to a numbered list, and add "More configurations." At the same time, attach theUCloud GPU configuration page。priceSource=reference ; when doing so, you must state that it is a historical reference price and that it will still be refreshed before creation.
  4. When the user selects "Other/More" or raises other requirements, run scripts/h3-onboard options --platform compshare --all --credential-profile <隔离名称>, and display the currently in-stock single-card specifications and real-time quotes. The list only represents what the platform currently sells, and does not mean that MiniMax H3 has been verified.
  5. After the user makes an explicit choice, run scripts/h3-onboard init --platform compshare --resource-option <配置ID> --credential-profile <隔离名称>. The default configuration on Windows is located at %APPDATA%\minimax-h3-cloud\config.json, and on macOS/Linux at ~/.config/minimax-h3-cloud/config.json; if that path already exists, read it first and confirm whether to reuse it, and do not overwrite it. This command does not accept an API Key, and it will not save an existing instance ID.
  6. Run for the isolated configuration scripts/h3-provision plan --config <隔离配置路径>First, internally inspect the region, image, cache, inventory, and real-time price of the selected specification. Do not switch to another configuration; if inventory is insufficient, return to step 3. Before generation, the confirmation shown to the user includes only the video mode, aspect ratio, duration, resolution, and estimated cost for this run, and does not display the image, cache, or internal evidence status. After obtaining user confirmation, run scripts/h3-provision create --yes --config <隔离配置路径>. A successful creation automatically saves the new instance ID.
  7. When a configuration already exists, first run scripts/h3-cloud status --config <配置路径>. Do not use developer default configurations, existing credential profiles, or existing instance IDs in a brand-new user acceptance.
  8. When the user prepares the prompt, remind him to specify the aspect ratio, duration, and resolution, and parse them according to the "Video Parameters" section. When all three are present, continue directly; when any are missing, ask only about the missing items. Write the picture prompt into a temporary UTF-8 text file, then pass it to --prompt-file; do not mix the aspect ratio, duration, and resolution into the workflow's picture prompt, and do not modify the baseline workflow file.
  9. Run scripts/h3-cloud generate --config <配置路径> --aspect-ratio <画幅> --duration-seconds <秒> --megapixels <MP> [--prompt-file <path>] [--first-frame-image <首帧绝对路径>] [--last-frame-image <尾帧绝对路径>] [--reference-image <参考图绝对路径>].... When no image is provided, it is text-to-video; passing only the first frame, only the last frame, or both first and last frames enters the corresponding keyframe mode respectively; reference image mode can pass 1–9 --reference-imagerepeatedly, and is mutually exclusive with the two keyframe parameters. Do not start the instance separately in advance; this command handles automatic start/stop, Runtime Prepare, reconnection recovery, and generation download. Use --preset。
  10. After the command ends, run scripts/h3-cloud status --config <配置路径>again, and you must confirm that the status is stopped。
  11. Use ffprobe to check the duration, resolution, video stream, and audio stream of the output MP4.
  12. Keep the final reply concise: first directly display the player using media Markdown with the local MP4 absolute path:![MiniMax H3 生成视频](<绝对路径>); immediately below the player, [下载 MP4](<同一个绝对路径>). Then list only in one line the actual duration, resolution, and that the GPU has stopped; expand other technical fields as needed, and do not require additional quality confirmation.

When the path contains spaces, the Markdown target must be wrapped in angle brackets. Do not return only a code block, a plain-text path, or a cloud temporary URL; the video must already be downloaded locally before it is presented.

If the generation command exits abnormally, check the manifest first: if prompt_id already exists, only the original task can be resumed, and it must not be regenerated; if there is no prompt_id , it means the submission has not yet been confirmed. Then query the instance status; if it is still running execute scripts/h3-cloud stop --yes, and report the original error, recovery result, and cleanup result. Do not automatically repeat paid generation when the cause is unknown.