SandTie PromptRig (H3)

Source repository on GitHub: https://github.com/sandpies/ComfyUI-SandTie-PromptRig · Full multi-hop chain engine: ComfyUI-Hand-Tie-Clips

SandTie PromptRig (H3)

Stop hand-typing 6-field MiniMax H3 walls of text. Drop SandTie PromptRig (H3) into any stock MiniMax H3 workflow (Ref2VA, T2VA, I2VA, FL2VA, L2VA, or p-h3-image), pick your cast & director dials in SandTie Studio, and drive either prompt at 0 VRAM or native positive + latent directly.

ComfyUI MiniMax H3 Zero VRAM CPU Mode MIT License


Why This Exists

MiniMax H3 wants strict six-field prompts (subject_definitions:, summary:, retention_analysis:, detailed_description:, overall_soundscape:, non_diegetic_music:), clean <Picture N> / <Subject N> / <Audio N> binding, zero negations outside <d> dialogue tags, and exact MM:SS.mmm beat pacing. Writing those by hand—or juggling LoadImage chains across every new workflow—kills flow.

SandTie PromptRig (H3) wraps the entire H3 authoring engine from ComfyUI-Hand-Tie-Clips into one plug-and-play ComfyUI V3 node:

  • Zero HTC dependency — install alone or side-by-side with ComfyUI-Hand-Tie-Clips with zero node, CSS, or HTTP route collisions.
  • Dynamic Autogrow sockets — starts with 1 ref_image_1 pin and expands up to ref_image_9 (ref_video_1..3, ref_audio_1..3) as you wire nodes, matching stock MiniMaxH3ReferenceToVideo.
  • Two drop-in modes on one node — run pure prompt (STRING) into any stock H3 node, or flip wire_mode to Direct H3 Conditioning (+ CLIP/VAE) to output native positive (CONDITIONING) + latent (LATENT) directly into BasicGuider & SamplerCustomAdvanced.
  • Persistent Reference Bible (S1..S9) — upload once, run a cached 2.5 MP Eye Scan, assign voice .wavs, and reuse cast across workflows without re-wiring LoadImage trees.

Two Ways to Wire into Stock MiniMax H3

Mode Set wire_mode to Wire IN Wire OUT Best When
1. Stock Prompt Drop-In Prompt Only -> Stock H3 Node Optional ref_image_1..9 / brief_in prompt → MiniMaxH3ReferenceToVideo.prompt / ImageToVideo.prompt Dropping straight into FastH3 / stock H3 templates (0 VRAM)
2. Direct H3 Conditioning Direct H3 Conditioning (+ CLIP/VAE) clip, vae, optional audio_vae + Bible or canvas refs positive → BasicGuider · latent → SamplerCustomAdvanced Casting from Reference Bible without LoadImage spaghetti

Ready-to-run graphs ship in example_workflows/:


Three Ways to Author

  1. ◩ CPU Tie (0 VRAM, Instant) — Deterministic multi-beat H3 assembler combining active dials, camera motion laws, weighted beat timelines (5–15 s), and Reference Bible descriptions. No LLM required.
  2. ⚡ Quick Tie (1-Pass) / Plan -> Write (Local LLM) — Calls any local OpenAI-compatible server (LM Studio, nInfer, llama.cpp, vLLM) with structured output, 2.5 MP Eye Scan vision context, and ✧ Double-Knot pass-2 continuity verification. Pick LM Studio / nInfer right next to the status pill (or click ⚙) inside Open SandTie Studio to set your URL, Model, and server path.
  3. ▶ Video -> Beats (30-FPS OpenCV Extractor) — Point at a reference .mp4, optionally trim IN/OUT, and extract motion cadence + camera trajectory + jump cuts directly into timed [Shot N] beats.

Inside SandTie Studio (⬒ Open SandTie Studio)

  • 19 Clean-Room Director Dials (2-Tier):
    • Primary 8: Cam (171), Light (71 options incl. Soft Natural Light & Soft Natural Window Light), Style (94), Place (127 incl. plain enclosed & window bedrooms + combat/dance plates), Wear (149), Action (178 incl. blow-for-blow ms fights), Stage (380), Speech (60).
    • Optics, Grade, Dance & Audio Drawer (11): Optics (94), Grade (44), FX (106, Delayed Mid / From Start), Dance / Pose (200 TikTok, K-Pop idol, waacking/crunk, duet & ensemble flows), Playbook (49), Cast Type (40), Genre (8), Era (17), Accent (30), Ambience (45 incl. Pure Music N/A), Music (52).
  • 114 Built-In Presets + Smart Roll: Domain-locked & aspect-safe (16:9 / 4:3 / 9:16) Surprise Roll, Match Dials from brief keywords, Compose Brief, 47 Clean Showcase Scripts, and ❝ Fresh Dialogue rewrites.
  • Built-In MiniMax H3 Rules: Zero non-<d> negations, automatic TTS cleanup (<d>[voice] "..."</d>), semicolon/speech stripping from overall_soundscape: & non_diegetic_music:, 12 H3 motion primitives, Open-End Range & Background Plate amplitude laws, and a live 7,000-char encoder counter.
  • All 6 MiniMax H3 Output Formats: Ref2VA / T2VA, I2VA (0.00 seconds anchor), FL2VA (0.00 + end-mark), L2VA (end-mark), Start Frame Prep (p-h3-image), and Storyboard Beats Only.

Quick Start

  1. Clone into ComfyUI/custom_nodes/:
    git clone https://github.com/sandpies/ComfyUI-SandTie-PromptRig.git
    
  2. Restart ComfyUI (requires ComfyUI v0.38.0+) and hard-refresh your browser (Ctrl+Shift+R).
  3. Drop SandTie PromptRig (H3) (under MiniMax H3) or load example_workflows/SandTie_Direct_H3_Conditioning.json.
  4. Optional local LLM writer: click ⬒ Open SandTie Studio → pick LM Studio / nInfer (or click ⚙ beside Start / Stop) to configure your URL, model, or server exe (or stay on Instant CPU Tie for 0-VRAM deterministic prompts).

Required Stock Models (for H3 Video Generation)

Component Recommended Model Folder
Diffusion Model minimax_h3_ref2va_int8_convrot (or FastH3 OmniRef) models/diffusion_models
Text Encoder qwen3vl_32b_minimax_h3_nvfp4_awq models/text_encoders
Video VAE minimax_h3_video_vae_int8_convrot models/vae
Audio VAE minimax_h3_audio_vae_fp32 models/vae

Credits & License

Built in ComfyUI-Hand-Tie-Clips (MIT). Architecture and prompting insights acknowledge LD Studio (original PromptMaster layout inspiration), wangweihong/comfyui-cinematic-gallery, and loopforge0/minimaxh3-shots-skills. See THIRD_PARTY_NOTICES.md.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for sandpies/ComfyUI-SandTie-PromptRig

Finetuned
(30)
this model