ComfyUI-MiniMaxH3-Easy

ComfyUI-MiniMaxH3-Easy

中文说明

A practical MiniMax H3 node suite for ComfyUI. It reduces the setup needed for text-to-video, image-to-video, first/last-frame generation, reference-to-video, digital humans, and long context-segment videos, while adding unified media management, visual prompt references, prompt optimization, and per-segment refinement.

Why use it?

• One main node for common generation modes: text, image, first/last frame, full reference, and digital human.
• Unified image, video, and audio management: upload, drag, paste, preview, reorder, replace, and remove media in one visual Media Loader. Video cards show their duration.
• Reference media directly in the prompt: type @ to insert an image, video, or audio reference without manually writing <Picture N> tags.
• Generate longer videos as connected segments: every segment keeps its own prompt and references while receiving visual or AV context from the previous segment.
• Built-in segment refinement: Pixel Resize, 3D Latent Upscale, and a Low VRAM Tile option.
• Still fully composable: samplers, LoRAs, attention patches, decoders, and save nodes remain ordinary ComfyUI connections.

Installation

First update ComfyUI to a recent version that includes the official MiniMax H3 nodes.

Install inside ComfyUI/custom_nodes:

git clone https://github.com/nkxx188/ComfyUI-MiniMaxH3-Easy.git

You can also search for ComfyUI-MiniMaxH3-Easy in ComfyUI Manager. Select the Nightly version during installation, because Nightly is the current up-to-date release; other published versions may lag behind this repository. Restart ComfyUI after installing or updating Python files.

Place the regular models in:

ComfyUI/models/diffusion_models/
ComfyUI/models/text_encoders/
ComfyUI/models/vae/

For a Latent Upscale refinement, place the matching H3 3D latent upscaler in:

ComfyUI/models/latent_upscale_models/

The Easy Loader can select FL2VA, Ref2VA, the text encoder, and both VAEs directly. Its filename whitelist and naming filters have been removed, so it shows every available file from the corresponding ComfyUI model directories; select the correct file for each role. Other native, community, or GGUF loaders can also be connected through MiniMax H3 Easy Model Adapter.

Quick start

The fastest route is to import workflow/1.MiniMax_H3_Easy.json, then:

1. Select the models in MiniMax H3 Easy Loader.
2. Select a mode in MiniMax H3 Easy.
3. Enter a prompt and connect any required image, video, or audio media.
4. Choose the resolution, aspect ratio, and duration.
5. Queue the workflow.

For a manual setup, the basic structure is:

Easy Loader → MiniMax H3 Easy → Easy Output → sample / decode / save

The main node’s Model output can continue through LoRAs, model patches, or a sampler. Connect H3 Context to MiniMax H3 Easy Output.

Generation modes

First/last-frame images are center-cropped when needed to fit the selected canvas. They are not stretched horizontally or vertically.

Both I2V or First/Last Frame and Reference-to-video automatically become text-to-video when no media is connected, so text generation is not tied to only one mode.

When both FL2VA and Ref2VA are configured, regular text/image generation prefers FL2VA and full-reference generation prefers Ref2VA. If only one transformer is configured, that model serves every mode.

Digital Human

Digital Human mode locks the single Media audio item into the generated result as its driving track; it is no longer treated as ordinary reference audio. Visual references may be images or videos. If no audio is supplied, the node automatically falls back to ordinary reference-to-video instead of failing.

Media and prompt editing

Four ways to provide media

Media Loader accepts images, videos, and audio dragged directly from the system file manager. You can also select only the Media Loader node and press Ctrl+V to paste supported media from the clipboard. Dragged or pasted files are placed into the appropriate media category automatically.

Audio cards include a compact play/pause control for quick checking. Video cards remain thumbnail previews.

Media Loader can hold a large shared library; the consuming Easy or Context Segments node applies the actual media limits. Of the four methods, only Media Loader caches decoded video references. After a video is decoded once, later generations can reuse it through ComfyUI’s node cache, reducing the loading and decoding time when the same reference video is used repeatedly. Replacing or modifying the file invalidates the cache automatically. Direct Media-port links, Media Bridge, and Media Splitter do not provide this decoded-video cache. It saves video preparation time, not the model’s sampling time.

Media Splitter is the reverse utility of Media Loader / Media Bridge: it expands one Media Bundle into standard IMAGE, VIDEO, and AUDIO outputs. Set the three counts and only that many output ports are shown; unused ports do not take up canvas space. This makes it useful for sending one shared media library into other ComfyUI workflows.

With direct Media links, drag from the port to an empty area to create a compatible media node, or click a virtual-link number to remove that item.

@ media references

Type @ in a reference or context prompt to select a connected image, video, or audio item. References can be displayed by index or filename and are converted to the H3 tags required at runtime.

Images, videos, and audio clips have independent numbering, so @Image1, @Video1, and @Audio1 can all exist together.

Dialogue blocks and raw view

Type # to create a dialogue block. It is converted to <d>...</d> at runtime. The controls at the bottom-right of the editor open prompt optimization and switch between the structured and raw prompt views.

The prompt can also be converted into a normal STRING input. When external text is connected, the built-in editor becomes read-only and the external string is used as the prompt.

Long videos with Context Segments

Context Segments divides a longer video into connected shots. Each segment keeps its own prompt, duration, and reference media while receiving continuity information from the previous segment.

The basic pipeline is:

Context Segments → Segment Sample → Segment Decode → first-pass video
                         ↓
                  Segment Refine → Segment Decode → refined video

Basic usage

1. Separate prompts with a standalone --- line. In the structured editor, type ~ to insert this divider automatically.
2. Enter matching durations in Segment seconds, for example 5,5,5.
3. Use @ to assign the required media to each segment.
4. Select a continuity mode and context-frame count.
5. Connect H3 Context to Segment Sample.

Regular first-pass workflow:

• 3.MiniMax_H3_Easy_Context_Segments.json

Optional per-segment control

If you only want one chain that generates every segment, use Segment Sample. To adjust seeds per segment, temporarily replace one segment’s prompt, or optionally rerun only the affected part later, use 5.MiniMax_H3_Easy_Context_Segments_Control.json. It connects the shared Context, Model, SAMPLER, and SIGMAS once through MiniMax H3 Easy Sample Setup, then chains multiple Segment Step nodes. The first run still generates the complete video; selective reruns are an extra capability.

The first Step receives Setup; later Steps only connect Previous segment, and the segment order is inferred from the chain. Each Step has its own seed, while Prompt override is optional. If it is not connected, the Step keeps using the segment prompt and @ media from Context Segments. The example contains 3 Steps; add or remove Steps as needed.

Segment prompts and reference media

All media connected to Context Segments forms one shared library, but it is not automatically used by every segment. A segment directly uses only the images, videos, or audio explicitly referenced with @ inside that segment’s prompt. References are not inherited from the previous segment, so repeat the @ reference in every segment that needs the same asset.

Media not referenced by the current segment is not sent into that segment’s generation, preventing references from different shots from interfering with one another. At runtime, the editor converts visible @ mentions into H3 tags and renumbers them for the subset used by that segment. A segment without a direct media reference can still receive continuity from the previous segment through the selected continuity mode.

Continuity modes

For speaker or singer identity, prefer an AV Prefix mode or supply a clear audio reference. RGB Guide mainly carries visual boundary information and cannot reliably lock voice timbre by itself.

Context Segments also supports Digital Human audio mode. Connect exactly one audio item and it will be sliced across the full timeline and used as the final complete soundtrack.

Per-segment refinement

MiniMax H3 Easy Segment Refine upscales and resamples one segment at a time while continuing to pass the refined result into the following segment. Prompts and reference media from different shots therefore remain separate.

• Pixel Resize: decode, resize, and re-encode; no latent upscaler model is required.
• Latent Upscale: use the built-in 3D latent upscaler instead of pixel resizing.
• Low VRAM Tile: spatially tile the current segment to trade more sampling time for lower VRAM use.
• Pixel Resize and Latent Upscale are now combined in one Context Segments workflow and placed in separate groups; the Latent Upscale group with the 3-step custom Sigma is enabled by default.
• Pixel Resize no longer loads a separate second-stage model; it reuses the first-pass model directly.

Example workflow (switch between the two groups as needed):

• 4.MiniMax_H3_Easy_Context_Segments_Refine.json

Segment Decode decodes one segment at a time into a temporary video file and returns a complete ComfyUI VIDEO with audio. It does not keep the full RGB timeline in memory.

Prompt optimization

Click ✦ at the bottom-right of the prompt editor to optimize the current prompt manually, or enable automatic optimization when the workflow runs. Supported providers include:

• OpenAI-compatible Chat Completions;
• OpenAI Responses;
• Gemini Native;
• Ollama;
• optional connected-media reading;
• mode-specific MiniMax H3 Prompt Guides;
• English and Chinese prompt modes, with a language-specific guide list. Chinese currently includes General, Dialogue, Action, and Advertisement.

Context-specific optimization

Segment seconds also tells the optimizer how many prompts to produce. For example, with 5,5,5, you can enter one undivided story or idea and click ✦ (or enable optimization on run). The optimizer will turn it into exactly 3 standalone prompts, separated automatically by ---.

• Whole sequence: understands the entire video in one request. It can turn one idea into a beginning, development, and ending, or improve an existing segmented script while preserving each segment’s intended action and order.
• Per segment: designed for an existing multi-segment draft. Segments are optimized separately with configurable concurrency; earlier original segments are supplied only as read-only continuity context, while only the current segment is rewritten.

This is more than generic prompt expansion. The optimizer adapts action density to each duration, carries the necessary subject appearance, setting, prop state, and audio cues into standalone prompts, and avoids planning phrases such as “Segment 2” or “continue from the previous clip.” When media reading is enabled, it also preserves or assigns @ references only where they are relevant. Generation still strictly follows the rule that each segment uses only the media referenced inside that segment.

Automatic optimization records prompts it has already processed. If the prompt, optimizer settings, and media are unchanged, it does not rewrite the prompt again on every run.

API settings are stored in prompt_optimizer.json inside the plugin directory. API keys are stored as plain text in that file, so do not publish or upload it.

Node reference

Limits and notes

• A regular reference generation can use up to 9 images, 3 videos, and 3 standalone audio clips, with 15 media items total.
• The Context Segments shared library accepts up to 27 images, 9 videos, and 9 audio clips. Each individual segment still follows the regular per-call reference limits.
• Digital Human mode accepts only one driving audio track.
• Media Loader caches decoded frames. Long or high-resolution reference videos can use substantial system RAM.
• The regular Segment Sample and Segment Refine nodes currently execute again on every queue, even when their seed and inputs are unchanged; the per-segment control workflow can use ComfyUI’s native cache for unchanged preceding Steps.
• Segment Decode requires FFmpeg. The project prefers imageio-ffmpeg and can also use FFmpeg from the system PATH.
• Prompt optimization is optional and is not required for generation.
• If an input field’s actual area becomes unusually wide or narrow, install ComfyUI-LegacyWidgetWidthFix and keep a Legacy Widget Width Fix node in the workflow for it to take effect; installing the plugin alone does not activate the fix. This is a ComfyUI frontend issue, not a bug in this node. Most users will not need this.

Additional information

• Both the classic ComfyUI canvas and Nodes 2.0 are supported.
• Chinese browser environments display Chinese labels; other environments default to English.
• Example workflows are available in workflow. Some workflows may require additional models, LoRAs, or custom nodes.

Acknowledgements

The context-continuity design was inspired by NikoDemon80/ComfyUI-H3-Motion-Context and its derivative project ethanfel/ComfyUI-MiniMaxH3-Contex-Loop; the code and node implementation were developed independently.

License and attribution

This project is released under the MIT License.

If your project reuses or adapts a substantial amount of this project’s code or major implementations, please credit the original author and mention ComfyUI-MiniMaxH3-Easy. Please do not present the reused portions as entirely independent original work. Minor references, ordinary use, or simply depending on this node do not carry this attribution request.