This project is a Human Assisted AI-driven storyboard-to-video pipeline idea designed to create a comedic, fan-made Dragon Ball Z series akin to That of TeamFourStar. By leveraging Models such as Nano Banana Pro and GPT 5 Image along with the new open source LTX-2 audio-visual model, we aim to transform static, stylized AI images and storyboards into cinematic video clips with synchronized sound.
The series follows "Bulack," a hyper-muscular, arrogant canyon-dweller who mistakenly believes he’s the strongest fighter in the universe. The comedy stems from the contrast between his massive ego and the unfiltered innocence behind the fact hes a black version of Bojack and a racist steriotype at best along with the completely oblivious Goku—leading to high-octane battles triggered by the most ridiculous misunderstandings.
- Image Generation: Create consistent 90s-style anime stills using specialized LoRAs to maintain the classic "VHS" aesthetic.
- Storyboarding: Use the Cinematic Script format to define camera angles, character movement, and lighting.
- Video Synthesis: Feed the images and prompts into LTX-2 to generate fluid motion and matching environmental audio in a single pass.
- Comedic Timing: Utilize the model's ability to sync audio-visual cues to emphasize "shaky cam" impacts and dramatic close-ups for comedic effect.
- Introduction to LTX-2
- Quick Start & Installation
- Required Models
- Available Pipelines
- Optimization Tips
- Prompting Guide
- Repository Structure
- Cinematic Script: Goku vs. Bulack
- Scene 1: The Encounter
- Scene 2: The Breaking Point
- Scene 3: The Explosive Opening
- Scene 4: Energy Clash
- Scene 5: The Victor
LTX-2 is the first DiT-based audio-video foundation model designed to generate synchronized video and audio within a single model. Developed by Lightricks, it brings together the core building blocks of modern video generation with open weights and a focus on practical, local execution.
- Joint Audio-Visual Generation: Generates visuals and sound simultaneously.
- High Fidelity: Supports up to 50 FPS and native 4K resolution (via upscalers).
- Flexible Modes: Includes both dev models and distilled versions for fast inference.
- Local Execution: Optimized for consumer hardware (e.g., RTX 4090/5090).
LTX-2 requires Python 3.12+ and CUDA 12.7+.
# Clone the repository
git clone https://github.com/Lightricks/LTX-2.git
cd LTX-2
# Set up the environment using uv (recommended)
uv sync --frozen
source .venv/bin/activate
Download the following assets from the LTX-2 HuggingFace repository:
| Component | Recommendation | Link |
|---|---|---|
| Model Checkpoint | ltx-2-19b-dev-fp8.safetensors |
Download |
| Spatial Upscaler | ltx-2-spatial-upscaler-x2-1.0 |
Download |
| Distilled LoRA | ltx-2-19b-distilled-lora-384 |
Download |
| Text Encoder | Gemma 3 (12B) | HuggingFace |
TI2VidTwoStagesPipeline: Production-quality text/image-to-video with 2x upsampling.TI2VidOneStagePipeline: Single-stage generation for quick prototyping.DistilledPipeline: Fastest inference with 8 predefined sigmas.ICLoraPipeline: Video-to-video and image-to-video transformations.KeyframeInterpolationPipeline: Interpolate between keyframe images.
- Use DistilledPipeline: Reduces steps to 8 (Stage 1) and 4 (Stage 2).
- Enable FP8: Use
--enable-fp8to lower memory footprint. - Attention Optimizations: Install
xFormersorFlash Attention 3. - Skip Cleanup: Disable memory cleanup between stages if VRAM allows.
Focus on detailed, chronological descriptions. Think like a cinematographer describing a shot list within 200 words.
Structure:
- Start with main action in a single sentence.
- Add specific movements and gestures.
- Describe character/object appearances precisely.
- Include environment, camera angles, and lighting.
| Package | Path | Description |
|---|---|---|
| ltx-core | packages/ltx-core/ |
Core model definitions and utilities. |
| ltx-pipelines | packages/ltx-pipelines/ |
High-level generation logic. |
| ltx-trainer | packages/ltx-trainer/ |
Tools for LoRA and fine-tuning. |
Visual Prompt: Cinematic anime style, 90s aesthetic (VHS grain, cel-shaded). Wide shot of a desolate, rocky canyon under a harsh sun. A massive, muscular warrior named Bulack with spiky red hair and pointed ears stands tall, arms crossed. Opposite him, Goku in a white and blue Gi stands relaxed. Dialogue:
- Bulack: "Sup. Names Bulack. I’m the muscle around these streets... and word has it there's a stronger fighter here in town today? That wouldn't be you, would it?"
- Goku (Close-up, chuckling): "Yeah? Depends on who’s asking, I guess!"
Visual Prompt: Medium shot of Bulack leaning in, his expression turning malicious. He says, "Big talk. I bet Chi-Chi won't be so cheerful when I bring back what's left of you." The camera cuts to a dramatic extreme close-up of Goku’s eyes; the pupils shrink as his face darkens with pure rage. The ground begins to tremble, and pebbles start to float around his feet.
Visual Prompt: High-octane action sequence. Goku screams "YOU’LL PAY!!" and lunges forward, becoming a blur of white light. He thrusts a powerful punch toward Bulack. Bulack catches the fist, and the impact creates a massive circular shockwave that shatters the surrounding rock pillars. The camera follows their movement in a fast-paced "shaky cam" style as they exchange a flurry of lightning-fast blows.
Visual Prompt: Dynamic mid-air shot. Goku flies backward, cup-shaping his hands to fire a brilliant blue Ki blast. Simultaneously, Bulack raises a hand, generating a massive, swirling crimson energy orb. The blue and red beams collide in the center of the screen, creating a blinding white light and electrical arcs that dance across the canyon walls. High-intensity particle effects and lens flares.
Visual Prompt: The smoke clears from a massive crater. Goku lies defeated on the ground, his Gi torn. The camera tilts up to reveal Bulack standing over him, untouched and towering. Bulack throws his head back in a menacing, guttural laugh, his red hair flowing in the wind. The sun sets behind him, casting a long, intimidating shadow over the fallen hero.