MiniMax H3 Prompt Generator (Portable)

A local desktop tool that generates properly formatted prompts for MiniMax Hailuo 3.0 (H3) video. Supports Reference-to-video + audio (RF2VA), Text-only, First frame, and First+Last frame modes. Runs completely offline after setup using your own Qwen GGUF model.

Features

• Dark UI with image drag-and-drop

• Live token streaming and cancel button

• Strict H3 template output

• Fully portable — no ComfyUI or system Python required

• Auto-opens the browser

Requirements

• Windows

• NVIDIA GPU with a recent driver

• CUDA runtime DLLs available (see Setup notes)

• A Qwen3.8-VL GGUF + matching mmproj file

Setup

1. Download and unzip the package.

2. Put your GGUF and mmproj files inside the models folder (or set the paths in the UI).

3. Double-click run.bat.

4. The browser opens automatically to http://localhost:7860.

GPU / CUDA notes

The package includes a CUDA-enabled llama-cpp-python wheel. For GPU offload to work you still need the CUDA runtime DLLs on your system.

• If you already have the CUDA toolkit installed (e.g. for ComfyUI), add its bin folder to the system PATH, or copy the cublas64_*.dll, cudart64_*.dll and related files into the python folder.

• If GPU isn’t used, the app falls back to CPU. Check the console on launch for layer assignment messages.

• There is no built-in setup window yet — the CUDA runtime must be present for GPU mode.

The wheel is detecting the GPU, but the CUDA runtime DLLs still aren’t being found when the app launches.

The simplest fix is to copy the needed DLLs into the python folder next to python.exe:

• From your CUDA bin folder (usually C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.x\bin or similar), copy:

  • cublas64_*.dll

  • cublasLt64_*.dll

  • cudart64_*.dll

  • nvJitLink_*.dll (if present)

Notes

• First load can take a minute while the model loads into VRAM.

• The tool does not include the model — you supply your own.

• GPU offload works if the CUDA wheel and runtime are present.