🎭 Welcome to MiniMax H3 Pure Prompt Forge — the ultimate, highly optimized 2-step text prompt engineering pipeline custom-built to maximize generation speed, asset stability, and video smoothness specifically on 8GB VRAM configurations (like RTX 3060, 4060, 5050).

Tired of Out-of-Memory (OOM) errors when loading heavy 16GB text encoders? This workflow completely reimagines the process for budget-friendly graphics cards by operating in an ultra-fast, pure text-to-text mode, bypasses safety filters natively via system prompts.

⭐ Why you should use this workflow:

- 🚀 Parallel Low-VRAM Pipeline: Splits complex physical action generation and audio-lipsync generation into two distinct, super-lightweight LLM steps. Your VRAM stays completely free from heavy vision layers, leaving max memory for the final video diffuser!

- 🎙️ Fully Automated Audio & Lipsync: No more manual coding or format messing. Type your prompt in ANY language, put spoken lines in quotes, and the engine automatically translates it to the strict single-quoted technical English syntax required by MiniMax H3.

- 🎵 Auto-Audio Dimming: Includes native tags that automatically keep diegetic music layered cleanly below the speech volume so voices never get muffled.

- 🔒 Fixed Character Appearance Block: Hardcodes facial and clothing stability markers directly into the text template so your character stays looking exactly like your reference image, even during intense dynamic dances or action sequences.

⚙️ Setup instructions & Canvas Connection:

1. Type your action and dialogue instructions into the User Prompt box in your native language (always put character spoken lines in quotation marks '').

2. Hit "Queue Prompt" — the pipeline will instantly output a perfectly compiled, censorship-free technical English prompt.

3. HOW TO CONNECT TO YOUR MINIMAX SAMPLER:

- Connect the final merged text output from this generator directly into the text/prompt input of your MiniMax H3 KSampler node.

- Load your character reference image into a standard "Load Image" node on your canvas.

- Connect that "Load Image" node directly into the "image" input of your MiniMax H3 KSampler (or its dedicated video diffuser loader).

This way, the video diffuser receives the clean face reference directly, while our lightweight text forge handles the heavy prompt and audio logic completely outside the VRAM, ensuring 100% stability and zero crashes on 8GB cards!

📦 Required Models:

- Text Encoder (CLIP): qwen3-vl-4b-heretic_nvfp4.safetensors (Strictly required for low VRAM text-only rewriter mode!)

- Diffusion Model: Any quantized H3 model (e.g., rzgar or pottokao format)