MiniMax H3 Novel Pipeline

MiniMax H3 Novel Pipeline

Turn novel chapters into organized character, location and object references, then prepare scene prompts for MiniMax H3. The plugin helps keep details consistent across chapters. It creates text and prompts; it does not generate images, audio or video.

Install

1. Install ComfyUI. Install LM Studio if using a local model.

2. Clone this repository into ComfyUI/custom_nodes/:

   git clone https://github.com/yuyuki/ComfyUI-Minimax_H3_Novel_Pipeline.git custom_nodes/minimax_h3_novel_pipeline
   

3. Install the plugin dependencies with the same Python environment used by ComfyUI:

   python -m pip install -r custom_nodes/minimax_h3_novel_pipeline/requirements.txt
   

4. Restart ComfyUI and refresh the browser.

First run

1. In ComfyUI Settings → MiniMax H3 Novel → Providers, enable LM Studio and/or OpenAI. Each card has its own Test and Save buttons; disabled cards collapse to their title and toggle. For LM Studio, load a supported model, start its API server and enter its URL and API key (use lm-studio if authentication is off). For OpenAI, enter your API key and model; the URL is fixed to https://api.openai.com/v1. OpenAI processing incurs API charges. Keys are masked, stored in ComfyUI settings, and excluded from workflows and outputs.
2. Add Provider Configuration and choose LM Studio or OpenAI. Only enabled providers with a saved API key and model appear. The node displays the saved model automatically and reads its API URL from settings. Qwen controls appear only for Qwen models in LM Studio. Test checks connectivity and model availability without generating text or saving drafts; Save applies the card. Fetch models in the provider card, select one and click Save before using the node. Restart ComfyUI and refresh the browser after updating this plugin.
3. Connect Select Chapters → Cinematic Chapter Adapter → Extract Chapter References → Consolidate References → Generate H3 Prompts. Share the configuration with each processing node and connect the chapter selection to Generate as well.
4. Choose your chapter files, then queue the workflow. The plugin saves results under ComfyUI/output/minimax_h3_novel/.

The chapter picker supports TXT, Markdown and searchable PDF files. Review the adapter's chapter scenes before extraction, or reload edited scenes with Load Cinematic Chapters. See examples for workflow details and saved-file loading.

Each processing node's thinking checkbox controls reasoning requests for Qwen, Gemma and Mistral in LM Studio. Saved configurations and stream logs reflect that setting. The loaded model and its chat template must support switching reasoning off; Gemma/Mistral requests fail if reasoning is returned while disabled. reasoning_chars reports the actual returned reasoning. GPT-OSS always reasons and reports thinking=true; its effort is set in LM Studio. Hosted OpenAI requests have no thinking control in this plugin. Character evidence auditing uses its separate audit_thinking checkbox.

Nodes

| Node | What it does | | --- | --- | | Provider Configuration | Shares the model and connection settings used by the AI processing nodes. | | Select Chapters | Picks chapter files or a folder to process. | | Cinematic Chapter Adapter | Turns chapter text into ordered, reviewable scene outlines. | | Load Cinematic Chapters | Loads saved scene outlines for another workflow run. | | Extract Chapter References | Finds characters, places, objects and story details in chapters. | | Load Chapter Catalogs | Loads saved chapter-by-chapter reference results. | | Consolidate References | Combines chapter references into a shared registry and creates image/audio briefs. | | Load Consolidated References | Loads a saved registry to continue without repeating consolidation. | | Generate H3 Prompts | Builds scene prompts and reference instructions for MiniMax H3. | | Spatial Continuity | Plans character, object, location and camera positions across scenes. | | Load Spatial Continuity | Loads and checks an edited spatial-continuity plan. | | Character Evidence Extractor | Collects quoted, chapter-linked evidence about character appearance. | | Load Character Evidence | Loads saved or manually edited character evidence. | | Consolidate Character Profiles | Groups evidence into source-linked character profiles. | | Load Character Profiles | Loads and validates saved character profiles. | | Character Visual Profile Generator | Creates permanent physical designs, then uses LM Studio to review possible duplicate identities and update the same files. | | Load Character Visual Profiles | Loads saved physical character profiles for Consolidate References. |

The visual generator first writes visual_designs.json, then checks possible duplicate identities against source evidence. Confirmed links update that file, character_visual_profiles.json, and the node output together. The upstream character profiles remain unchanged. Uncertain cases and conflicting edited defaults are listed in character_visual_identity_review.json; existing manual identity links are preserved. Review requests use the generator's .cache/ folder and its optional token, context and retry controls.

Major changes by release

Highlights for repository releases; see the full changelog for details.

| Release | Highlights | | --- | --- | | 2.0.0 | Adds a source-linked character evidence and profile workflow, including visual profiles and loaders. | | 1.10.x | Improves scene pacing, reference selection, image prompt exports and publishing checks. | | 1.9.0 | Adds Qwen/Mistral model selection, model profiles and stronger prompt validation. | | 1.4.0 | Adds registry summaries and improved visual design handling. | | 1.0.0 | Establishes the chapter selection and extract, consolidate, generate workflow. |

More

• Example workflow and wiring guide
• Full changelog
• License