🚀 V2 INT8 — Latest Update

Minimax-h3_Singularity V2 INT8 builds on V1 with expanded action coverage, stronger prompt adherence, a fix for audio reference conditioning, and enhanced visual style rendering.

✨ What's New in V2

• ⚔️ Expanded Action Coverage: Further optimized on top of V1 to support a wider range of action scenes and more varied motion sequences, extending the model's versatility across dynamic video generation scenarios.
• 📝 Stronger Prompt Adherence: V2 follows prompt instructions more closely, improving control over subjects, actions, environments, and visual styles. We recommend keeping prompts within approximately 5,000 characters, using clear, focused descriptions and avoiding excessive or repetitive instructions.
• 🎵 Audio Reference Fix: Fixed the audio reference conditioning issue. Audio references can now be used to help adjust the character's color tone.
• 🎨 Enhanced Visual Styles: Strengthened visual style rendering, with strong results across photorealistic, East Asian realism, 3D CG, and cinematic CG scenes. For best results, explicitly include the desired style keywords in your prompt.
• 📦 INT8 Release: The updated V2 model is available in INT8 format.

💡 Recommended Prompting for V2

Clearly describe the subject, action, environment, camera movement, and intended visual style. Keep instructions consistent and prioritize the details that matter most.

Include style keywords that match your desired result. Examples:

Recommended prompt length: within approximately 5,000 characters. This is a practical recommendation, not a target to fill. Shorter, well-structured prompts are welcome; avoid unnecessary details, repetition, and conflicting style instructions.

---

Minimax-h3_Singularity

📖 Model Overview

Minimax-h3_Singularity is a comprehensive fine-tuned fusion model designed to enhance the capabilities of MiniMax-H3. As a versatile multimodal video generation model, it supports Text-to-Video (T2V), Image-to-Video (I2V), Reference-to-Video (Ref2V), and Video-to-Video (V2V) workflows within ComfyUI.

Built upon a strategic fusion of key checkpoints, including ref, fl, and b25-49, the model underwent extensive high-step fine-tuning. To preserve the original model's foundational strengths and broad generalization while addressing artifacts introduced during training, we spent 3 full days on precise model pruning and weight optimization. The result is a clean, sharp, and highly dynamic video generation model.

The latest V2 INT8 release builds on this foundation with broader action coverage, improved prompt adherence, corrected audio reference conditioning, and stronger visual style rendering.

---

✨ Key Improvements & Features

• 🎬 HDR Image Quality & Blur Reduction: Fine-tuned on high-dynamic-range (HDR) video datasets to enhance visual clarity and reduce unwanted motion blur during high-speed action.
• 👤 Distant Face Restoration: Reduces facial distortion, blurriness, and structural collapse in medium and long shots.
• 🎨 Clean & Natural Aesthetic: Reduces excessive skin shine and unnatural glossy textures for more natural lighting and realistic materials.
• ⚔️ Enhanced Dynamic Motion: Improves motion fluidity and physical impact, including complex sequences such as sword fighting, martial arts, and melee combat. V2 expands support to a wider range of action scenarios.
• 🌌 VFX & Fantasy Effects: Optimized for fantasy spellcasting, particle auras, and magical combat visual effects.
• 🎭 Expressive Facial Dynamics: Captures subtle facial expressions and emotional nuances more vividly.
• 📹 Cinematography & Camera Control: Improves responsiveness to camera instructions, including pans, tilts, zooms, and tracking shots.
• 📝 Improved Prompt Control in V2: Stronger adherence to prompt instructions enables more precise control over scene content and presentation.
• 🎵 Corrected Audio Reference Conditioning in V2: Resolves the audio reference issue and supports using audio references to help adjust the character's color tone.
• 🖼️ Stronger Style Rendering in V2: Delivers improved results in photorealistic, East Asian realism, 3D CG, and cinematic CG styles when the intended aesthetic is specified in the prompt.
• 🛡️ Preserved Base Capabilities: Designed to retain MiniMax-H3's core multimodal generation capabilities, style adaptability, and broad generalization.

---

🎬 Showcase

---

💡 Usage Guide

Multimodal Pipeline Support

The model supports the following workflows within ComfyUI:

• Text-to-Video (T2V)
• Image-to-Video (I2V)
• Reference-to-Video (Ref2V)
• Video-to-Video (V2V)

📝 Prompt Recommendations for V2

For best results:

• Keep prompts within approximately 5,000 characters.
• Use clear descriptions of the subject, action, setting, lighting, and camera movement.
• Explicitly include the intended visual style, such as photorealistic, East Asian realism, 3D CG, or cinematic CG.
• Avoid repetitive descriptions, unnecessary instructions, and conflicting visual styles.
• For action scenes, describe the movement and sequence of events clearly.

🎨 Example Style Prompts

Photorealistic

Photorealistic style, natural lighting, realistic skin texture, detailed environmental materials, fluid character movement, cinematic camera tracking.

East Asian Realism

East Asian realism, a traditional courtyard, realistic fabric textures, soft natural lighting, detailed architecture, fluid martial arts movements, cinematic composition.

3D CG

3D CG style, detailed character materials, expressive facial animation, fluid action, depth and dimensionality, polished 3D rendering.

Cinematic CG

Cinematic CG style, dramatic cinematic lighting, detailed environments, dynamic action choreography, atmospheric effects, film-quality rendering.

🎵 Audio Reference

V2 fixes the audio reference conditioning issue. You can use audio references to help adjust the character's color tone as part of your generation workflow.

🚀 Recommended Acceleration LoRA

For high-speed generation, the recommended acceleration LoRA is:

• minimax_h3_ref2v_turbo_4step_v0.1 — enables 4-step fast inference.

🌐 Online Interactive Demo

Test the model directly in your browser without a local GPU setup:

👉 Try it on RunningHub Workflows

---

🙏 Acknowledgements

Special thanks to the MiniMax open-source team for creating and releasing the powerful MiniMax-H3 multimodal video model and providing a solid foundation for the open-source community! 🤝

---

🤝 Community & Commercial Inquiries

Connect with us for tutorials, community discussions, workflow sharing, or commercial collaborations:

• YouTube Channel: AIGC-Singularity
• Bilibili Channel: AIGC-Singularity Space
• QQ Group 1: 1058747239 (Request to join)
• QQ Group 2: 1072010342 (Request to join)
• Business Inquiries (WeChat): aigctyd
• Email: a592991299@gmail.com