Extensions/ComfyUI Vergora Minimax H3 Video
ComfyUI Extension

ComfyUI Vergora Minimax H3 Video

ComfyUI node for minimax-h3 video generation through the Vergora API

By vergora-megatron·Created a day ago·Updated about 20 hours ago· 0
vergora-megatron/ComfyUI-Vergora-Minimax-h3
Nodes—
On cloudLocal install
Stars0
Updatedabout 20 hours ago
Readme

ComfyUI Vergora Minimax H3 Video

A ComfyUI custom node that generates video with minimax-h3 through the Vergora API. It is a separate package from ComfyUI-Vergora-Wan3 because the two models do not share a contract: minimax-h3 is image-driven, accepts only three ratios, and only one resolution is known to work. Both packs can be installed side by side.

  • Model: minimax-h3, fixed by the package.
  • Input: one or more reference images (an IMAGE batch or URLs) plus a prompt.
  • Verified on: ComfyUI 0.40.0 (frontend 1.55.17), Python 3.13, Windows 11.

What this model accepts

These limits come from testing the channel behind minimax-h3 (2026-10-15). They are baked into the widgets, so you cannot accidentally pick a value the upstream will reject.

| Setting | Accepted | Notes | | --- | --- | --- | | Reference image | required | Text-only jobs fail with 缺少必填参数：ref_image_0; the node refuses them locally | | resolution | 1080p | 720p is rejected with invalid_resolution, so it is not offered | | ratio | 16:9, 9:16, 1:1 | There is no adaptive on this channel | | duration | 2 to 30 seconds, or -1 | -1 asks the service to pick the length | | audio, prompt_extend, seed | as in the W3.0 pack | seed = -1 means "let the service choose" |

The reference image is connected as an IMAGE; a batch of up to 10 is fine. It is deliberately the only image input: sending a URL instead would make the upstream fetch that URL, which fails whenever the link stops resolving (4002 Failed to download file), and a second image input would also make ComfyUI print the validation message once per input. Use a LoadImage node or the output of an image generation node. reference_video_urls and reference_audio_urls exist because the channel's request body has videos / audios fields, but they have not been exercised yet.

Requirements

  • ComfyUI (developed against 0.40.0)
  • A Vergora API key
  • No extra pip packages: the HTTP layer uses requests when present and falls back to the standard library otherwise.

Installation

  1. Copy this folder into ComfyUI's custom_nodes directory.

  2. Create the configuration from the template and add your key:

    • Windows: copy config.ini.example config.ini
    • macOS / Linux: cp config.ini.example config.ini
    [API]
BASE_URL = https://api.vergora.ai/v1
KEY = sk-your-key-here

    Every setting is documented inside config.ini.example. The environment variables VERGORA_BASE_URL and VERGORA_API_KEY override the file.

  3. Restart ComfyUI and add vergora minimax-h3 视频生成（参考图） from the vergora category (search for minimax).

config.ini is gitignored, so your key never reaches the repository.

Node reference

Node id: VergoraMinimaxH3Video Category: vergora

| Input | Type | Default | Notes | | --- | --- | --- | --- | | prompt | STRING (multiline) | empty | What should happen in the clip; address inputs as Image 1, Video 1 | | resolution | COMBO | 1080p | The only value the channel accepts today | | ratio | COMBO | 16:9 | 16:9, 9:16 or 1:1 | | duration | INT | 5 | -1 for smart duration, otherwise 2 to 30 seconds | | audio | BOOLEAN | true | Whether the clip carries an audio track | | prompt_extend | BOOLEAN | true | Let upstream rewrite the prompt | | seed | INT | -1 | -1 means "do not send a seed" | | reference_images | IMAGE (optional) | - | Required in practice: the reference image(s); a batch of up to 10 works | | reference_video_urls | STRING (multiline, optional) | empty | Reference videos, one URL per line (untested) | | reference_audio_urls | STRING (multiline, optional) | empty | Reference audio, one URL per line (untested) | | extra_json | STRING (multiline, optional) | empty | Extra request fields, merged last |

Outputs: video_path (local .mp4), task_id, video_url. The clip also appears in ComfyUI's result panel, and the node registers as an output node.

Usage

Reference image to video. Connect an IMAGE (a LoadImage node, an image-generation node, or a batch) to reference_images, write what should happen in prompt, and press Run. The console prints the submit call and one status line per poll, including the upstream's own progress percentage.

Ready-made workflow. Load examples/minimax-h3-reference.json (Workflow → Open): it wires a LoadImage node into reference_images and already selects 1080p / 16:9. Pick any image of your own and press Run.

Cost

minimax-h3 is billed per request, not per second, so changing resolution or duration does not change the price of a successful job. The Vergora catalog listed it at 1.365 USD per request (70% discount) when this package was written; confirm the current rate in your console. Failed jobs are not billed.

Troubleshooting

| Symptom | Cause | | --- | --- | | minimax-h3 必须提供参考图 (shown as invalid input, exactly one line) | Nothing is connected to reference_images. Wire a LoadImage node or an image generation node into it | | 没有配置 API key | KEY is empty in config.ini and VERGORA_API_KEY is not set | | 503 No available channel for model minimax-h3 | This key has no channel for the model; ask the platform to enable it | | 402 Insufficient balance | The request costs more credits than the account holds; the message states both amounts | | invalid_resolution in a task failure | The channel rejected the resolution. Only 1080p is offered; clear any resolution override in extra_json | | 4002 Failed to download file | The upstream could not fetch a reference URL. Use an IMAGE input (sent as Base64) or a stable public URL | | 413 Request Entity Too Large | Inline images exceeded the gateway body limit. The node budgets and compresses automatically; otherwise ask the platform to raise client_max_body_size | | SSLError / UNEXPECTED_EOF_WHILE_READING | A dropped TLS connection, usually a local proxy or VPN. Submits are not retried automatically (a duplicate could be billed twice) — press Run again | | The job completes but the file is 0 bytes | The CDN dropped the transfer; the node verifies and retries three times, then reports the video URL so you can fetch it by hand |

Development

python test_offline.py runs the whole flow (submit → poll → download → preview) plus the input rules against a stub HTTP layer. It needs no network, no key and spends nothing.

License

MIT. See LICENSE.

中文说明

ComfyUI Vergora minimax-h3 视频节点：通过 Vergora API 调用 minimax-h3 生成视频。它与 Wan 3.0 不是同一套契约，所以单独成包（Wan 3.0 见 ComfyUI-Vergora-Wan3），两个包可以同时安装。

这个模型的硬约束（2026-10-15 实测，已固化到控件里）

  • 必须给参考图：纯文本会失败（上游报 缺少必填参数：ref_image_0），节点会在提交前拦下并提示
  • 分辨率只有 1080p：720p 会被上游拒绝（invalid_resolution），所以下拉里不提供
  • 画幅只有 16:9 / 9:16 / 1:1：这个通道没有 adaptive
  • 时长 2–30 秒或 -1（智能时长）；audio、prompt_extend、seed 与 Wan3 包一致

安装：整个文件夹放进 custom_nodes → 把 config.ini.example 复制成 config.ini 并填上 KEY → 重启 ComfyUI → 在 vergora 分类下添加「vergora minimax-h3 视频生成（参考图）」。 也可以用环境变量 VERGORA_API_KEY；config.ini 已被 gitignore，不会提交密钥。

用法：把图片接到 reference_images（支持一批，最多 10 张），填好 prompt，点运行； 参考图只能用 IMAGE 接进来（LoadImage 或出图节点都可以，一批最多 10 张）。这里刻意不提供 "填图片链接"的输入：一是链接要由上游自己去抓，链接一失效就会以 4002 Failed to download file 失败；二是多一个图片输入会让 ComfyUI 把校验提示重复打印多次。想用现成图片文件，就用 LoadImage。 开箱即用的工作流在 examples/minimax-h3-reference.json，加载后换一张自己的图即可。

计费：按请求计费（不是按秒），改分辨率或时长都不影响价格；撰写本包时目录价 1.365 美元/次 （70% 折扣），以控制台当日价格为准。失败任务不计费。

自测：python test_offline.py，用假 HTTP 层跑通全流程与输入校验，不联网、不花钱、不需要密钥。