🎬 Director Vance | MiniMax H3 Video Prompt Studio 🎬

> "Got a scene in your head? Let's bring it to life frame by frame."

---

### 💡 What is this Bot?

This is an experimental utility/RP hybrid bot designed to generate optimized, highly descriptive video prompts for MiniMax H3 (H3 Omni) — the general-purpose multimodal video model capable of generating up to 2K resolution, 24fps video with native stereo audio.

Whether you're crafting cinematic movie clips, anime scenes, sci-fi sequences, or sensual atmospheric shots, Vance acts as your personal AI Cinematographer. He translates your rough ideas into production-ready prompts structured specifically for H3 Omni's multimodal architecture (including motion, camera work, lighting, and audio layers).

---

🧪 Public Test Chat Logs (LLM Model Comparisons):

Compare how Vance performs across different LLM setups in real testing sessions.

(Note: Since Janitor AI cannot directly embed videos, full generated videos are shared on my Discord. However, I have attached the *First Frame images** of the generated videos below! )*

- 🔹 Janitor LLM (Default): [View Test Chat Log](https://janitorai.com/ja/chats/public/test-1-janitor-llm-default-model)

📸 Generated Video First Frames:

* SCENE 1: Midnight Cyberpunk Cafe

* SCENE 2: Deep Space Battleship Combat

* SCENE 3: Cozy Bedroom on a Rainy Day

The image shows a real woman resting in bed in a camera-captured bedroom scene, with no explicit or otherwise restricted content.

- 🔹 GLM 5.1: [View Test Chat Log](https://janitorai.com/ja/chats/public/test-2-glm-5-1-proxy)

📸 Generated Video First Frames:

* SCENE 1: Midnight Cyberpunk Cafe

The image shows a real camera-captured photograph of a man seated in a café or diner.

* SCENE 2: Deep Space Battleship Combat

* SCENE 3: Cozy Bedroom on a Rainy Day

The image is a realistic photograph showing an adult woman and an adult man resting together in bed, with no explicit or otherwise prohibited content visible.

- 🔹 DeepSeek V4 Flash (0731): [View Test Chat Log](https://janitorai.com/ja/chats/public/test-3-deepseek-v4-flash-proxy)

📸 Generated Video First Frames:

* SCENE 1: Midnight Cyberpunk Cafe

* SCENE 2: Deep Space Battleship Combat

* SCENE 3: Cozy Bedroom on a Rainy Day

The image shows a real woman sleeping in a bedroom without explicit or otherwise restricted content.

💡 My Quick Impressions on Each Model:

- Janitor LLM (Default): Surprisingly well-balanced!

- GLM 5.1: Solid, safe, and reliable.

- DeepSeek V4 Flash: Excellent at generating audio/sound prompts!

---

💬 See Real Generated Videos & Community Discord:

Want to see the actual video results generated from Vance's prompts using a local MiniMax H3 setup? Or have requests/ideas for video generation?

👉 [Click Here to Join my Discord Server](https://discord.gg/8CtUCdVmT9)

(Please note: English is not my native language and I barely understand it. I am using AI translation tools to write all of this text. I sincerely apologize in advance if there are any misunderstandings, awkward wording, or misinterpretations in our communication!)

(Feel free to drop requests or check out rendered showcases! Instant generation is not guaranteed, but feel free to hangout!)

💻 Local Video Generation Specs (Discord Showcases):

Showcase videos shared on Discord are rendered using a local H3 First/Last setup with the following parameters:

- Model: H3 First/Last Model (Local)

- Mode: Text-Only

- Resolution & Aspect Ratio: 480P | 16:9

- Duration: 5 Seconds

- Text Encoder: NVFP4 QWQ

- Inference Steps: 8 | Guidance Scale: 1 | Seed: 130000756

- PC Specs: NVIDIA GeForce RTX 3060 12GB (Approx. 15 minutes per video)

---

⚠️ PLEASE NOTE (Experimental Disclaimer & Settings):

* Experimental Bot: This bot is built primarily for experimental and testing purposes. Please keep in mind that it may not always yield the absolute perfect or optimal prompt for every scenario.

* Prevent Response Truncation: To prevent the generated code block from cutting off mid-sentence, please make sure to set your "Max New Tokens" (Generation Max Tokens) setting to 0 (unlimited) in your Generation Settings or Proxy setup.

---

### 🛠️ How to Use

1. Give him a concept: Simply tell Vance what you want to see (e.g., "A cybernetic street racer in Tokyo under neon rain" or "A fantasy witch brewing a glowing potion").

2. Review the Breakdown: Vance will analyze the scene, suggest cinematic camera angles and soundscapes, and deliver a formatted H3 Omni prompt in a code block.

3. Copy & Generate: Copy the code block directly into Hailuo AI / MiniMax H3 / ComfyUI / fal.ai and hit generate!

4. Refine: Ask him to adjust lighting, camera speed, resolution aspect ratios, or audio cues anytime.

---

### ⚙️ Notice for Proxy & External Model Users (DeepSeek, Gemma, GLM, etc.)

If you are using an external model via Proxy and notice that the bot is prioritizing lengthy roleplay over generating the complete prompt code block, please copy and paste the following snippet into your Custom Prompt / System Prompt settings to get the best result:

```text

[SYSTEM OVERRIDE: CODE GENERATION & UTILITY FIRST]

Priority Rule: If the active bot is generating code, scripts, or formatted prompts (such as Markdown code blocks), OVERRIDE all default story-writing and narrative expansion rules.

1. Limit conversational roleplay and asterisk actions to a maximum of 2 short sentences.

2. Allocate at least 80% of the output length to delivering complete, non-truncated Markdown code blocks.

3. Never summarize, shorten, or omit the generated code/prompt.

4. Always prioritize functional utility and exact structural formatting over narrative roleplay.