Usage Guide

Learn how to get the best results from LTX Video Generator.

Table of contents

  1. Basic Workflow
    1. 1. Write Your Prompt
    2. 2. Choose a Preset
    3. MiniMax H3 variants
    4. 3. Generate
    5. 4. View Results
  2. Gemma Prompt Enhancement
  3. Image-to-Video
  4. Adding Audio
    1. Voiceover / Narration
    2. Background Music
    3. Music Genre Categories
    4. Audio from History
  5. Writing Effective Prompts (LTX-2.5)
    1. Six ingredients
    2. Structure
    3. Leave enhancement off when dense
  6. Using the Queue
    1. Add Multiple Generations
    2. Batch Variations
    3. Queue Management
  7. History Features
    1. Browse Videos
    2. Video Details
    3. Actions
  8. Tips for Best Results
    1. Start Small
    2. Use Negative Prompts
    3. Reproducible Results
    4. Memory Management

Basic Workflow

1. Write Your Prompt

Enter a descriptive text prompt. For LTX-2.5, write one flowing present-tense paragraph (shot, lighting, action, character, camera, audio). Spoken lines only in quotes.

Good prompt:

Handheld medium close-up, cool overcast light on a street corner. A busker in his late 40s — weathered face, grey stubble, battered acoustic guitar — stops mid-riff and looks past the camera. Soft traffic hum. He speaks in a low urgent whisper, “They’re coming.” He pauses and holds the silence.

Less effective:

“busker scary” / Beat. / JUMP CUT 1 (2.5–5s): …

2. Choose a Preset

Select a preset from the dropdown to quickly configure parameters:

Preset Resolution Frames Steps Best For
Quick Preview 512×320 49 15 Quick tests
Standard 768×512 121 30 Balanced quality
High Quality 768×512 121 40 Best results
Portrait 512×768 97 30 Vertical videos
Square 512×512 97 30 Social media

The LTX-2 Unified model uses a 2-stage generation pipeline with built-in audio. ltx-2-mlx distilled (LTX-2.5 Distilled / Q8 DiT and the 12GB 2.3 pack) is a single 8-step pass. ltx25_dev is half-res dev+CFG then distilled-LoRA refine. MiniMax H3 runs the native h3 binary (24 fps, frames snap to 5+17n).

MiniMax H3 variants

All three stay on Metal h3.c. No MLX-H3, ANE, GGUF, or ComfyUI runtime.

model_id Weights Disk RAM Steps Notes
minimax_h3 Official MiniMaxAI BF16 ~144GB 32GB+ 4 / 20 / 50 presets Stock antirez/h3.c
minimax_h3_int8 Comfy-Org int8 DiT + MiniMaxAI TE/VAE ~92GB (~20GB extra if BF16 is cached) 24GB+ Same presets h3.c-int8 (GPU ConvRot de-rot). Slight quality drop.
minimax_h3_turbo Official BF16 + folded Turbo v4 LoRA ~144GB + ~3GB CoW 32GB+ Fixed 6 First generate bakes the LoRA. Do not combine with reuse.

3. Generate

Click the Generate button. Progress shows:

  • Model loading status
  • Stage 1 and Stage 2 denoising progress
  • Video encoding and saving

4. View Results

  • Queue sidebar shows real-time progress
  • History tab displays all generated videos with thumbnails
  • Videos save to your configured output directory

Gemma Prompt Enhancement

Improve results by having Gemma rewrite your prompt with vivid details:

  1. Go to Settings > Generation and enable Enable Gemma Prompt Enhancement
  2. In the prompt view, expand Prompt Enhancement (Gemma)
  3. Optionally enable Use uncensored enhancer to avoid content filters (first run downloads ~7GB)
  4. Click Preview enhanced prompt to see the rewritten prompt before generating
  5. Generate as usual—the enhanced prompt is used automatically

If enhancement returns empty (e.g. safety filter), the app auto-retries with filtered words replaced, then merges originals back.

Image-to-Video

You can animate images into videos:

  1. On the main Generate screen (prompt column), expand the Image to Video disclosure section — it sits below Prompt Enhancement and above Negative Prompt.
  2. Click Select Source Image… and pick an image file; it becomes the conditioned first frame.
  3. Optionally adjust Image Strength (1.0 = full influence, lower = more motion freedom).
  4. Write a prompt that describes the motion; then generate as usual.

Adding Audio

Voiceover / Narration

Add text-to-speech voiceover to your videos:

  1. Expand the Voiceover / Narration section
  2. Choose your source:
    • MLX Audio (Local) - Free, runs on-device, good quality
    • ElevenLabs (Cloud) - High quality, requires API key
  3. Select a voice from the dropdown (10 voices for MLX, 9 for ElevenLabs)
  4. Enter your narration text
  5. Generate your video - audio will be added automatically

You can also add audio later by right-clicking any video thumbnail in the Video Archive.

Background Music

Add AI-generated instrumental music (requires ElevenLabs API key):

  1. Expand the Background Music section
  2. Toggle Generate background music on
  3. Choose a genre from the dropdown:
    • Organized into 9 categories with 54 total presets
    • Preview the prompt by hovering over the selection
  4. Music automatically matches your video length

Music Genre Categories

Category Genres
Electronic EDM, House, Techno, Ambient, Chillwave, Synthwave, D&B, Trance
Hip-Hop/R&B Trap, Lo-Fi, Boom Bap, Slow R&B, Modern R&B, Soul
Rock Classic, Alternative, Indie, Metal, Punk, Acoustic
Pop Modern, Indie, Dance, Acoustic
Jazz/Blues Smooth, Bebop, Lounge, Electric Blues, Acoustic Blues
Classical/Cinematic Orchestral, Piano, Chamber, Epic, Tense, Uplifting
World Latin, Reggae, Afrobeat, Middle Eastern, Asian
Country/Folk Modern, Classic, Acoustic Folk, Indie Folk
Functional Corporate, Motivational, Relaxing, Suspense, Action, Romantic, Happy, Sad, Dramatic, Mystery

Audio from History

Add audio to previously generated videos:

  1. Go to Video Archive
  2. Right-click any video thumbnail
  3. Select Add Audio (or Replace Audio if it already has audio)
  4. Choose from three tabs:
    • Voiceover - Add narration only
    • Music - Add background music only
    • Both - Add voiceover and music together

When combining voiceover and music, the music is automatically ducked to 20% volume so the voice remains clear.

Writing Effective Prompts (LTX-2.5)

Follow the official LTX-2.5 guides: prompt guide · docs. Copy-paste examples: EXAMPLES.md.

Six ingredients

  1. Shot — scale / lens / genre look
  2. Scene — lighting, palette, atmosphere
  3. Action — present-tense verbs that move
  4. Characters — age, clothes, physical emotion cues (not abstract labels)
  5. Camera — how and when it moves
  6. Audio — ambience, music, speech

Structure

  • Single take: one flowing paragraph, ~4–8 sentences, present tense.
  • Dialogue: only spoken words in "quotes"; name language / accent / delivery.
  • Pauses: write as action (he pauses, holds the silence). Avoid bare Beat. — the model may speak it.
  • Multishot: one chronological paragraph; name cuts in prose (A hard cut transitions to…); re-ID subjects and state audio continuity. Do not use START FRAME / numbered JUMP CUT lists.
  • Mac / Distilled–Dev: prefer a single continuous take for dialogue + face; native multishot / Prompt Relay / DFR is incomplete on ltx-2-mlx.

Leave enhancement off when dense

A well-formed 2.5 prompt often needs no rewrite. Gemma / --enhance-prompt helps short or foreign-model prompts; leave it off when you already wrote the full cinematic paragraph.

Using the Queue

Add Multiple Generations

  1. Write your prompt
  2. Click Add to Queue (instead of Generate)
  3. Modify the prompt or parameters
  4. Add more to the queue
  5. Videos generate one after another

Batch Variations

Click the batch menu (stack icon) to:

  • Generate 3 variations
  • Generate 5 variations
  • Each uses a random seed for different results

Queue Management

  • Cancel the current generation with the X button
  • Remove pending items from the queue
  • Clear the entire queue with the Clear button

History Features

Browse Videos

  • Thumbnails show a frame from each video
  • Sort by newest, oldest, or prompt alphabetically
  • Search prompts to find specific videos

Video Details

Click a video to see:

  • Full video preview (loops automatically)
  • Original prompt
  • All generation parameters
  • Timestamp and generation duration
  • Seed value for reproducibility

Actions

  • Show in Finder - Reveal the video file
  • Share - Share via macOS share sheet
  • Reuse Prompt - Copy prompt back to input
  • Delete - Remove video

Tips for Best Results

Start Small

  • Use Quick Preview preset first
  • Iterate on prompts quickly
  • Only increase quality for final renders

Use Negative Prompts

Click the disclosure arrow to add negative prompts:

worst quality, blurry, jittery, distorted, watermark

Reproducible Results

  • Note the seed value of good generations
  • Enter the same seed to reproduce results
  • Useful for making variations with slight prompt changes

Memory Management

  • Higher resolutions use more memory
  • Close other apps if you encounter issues
  • 32GB RAM minimum, 64GB recommended

Back to top

Copyright © 2026 James Campbell. Distributed under the MIT License.

This site uses Just the Docs, a documentation theme for Jekyll.