LTX-2.5: A Complete Guide to Lightricks’ Audio-Video AI Model

Overview

LTX-2.5 is a DiT-based audio-video foundation model built by Lightricks that generates synchronized video and audio in a single unified pass. The model uses an asymmetric dual-stream transformer architecture with a 22B-parameter distilled transformer for video generation and separate audio encoding, coupled through bidirectional audio-video cross-attention layers with modality-aware classifier-free guidance. It accepts text prompts, images (for image-to-video), or existing video clips as inputs and outputs MP4 files containing both video and synchronized audio. Key innovation in the 2.5 release includes native multishot generation—generating multiple connected scenes with consistent character identity, environment, lighting, and voice across cuts in a single pass—along with a new diffusion video decoder replacing VAE reconstruction for sharper details and improved motion. The model uses a custom Gemma 4 12B text encoder for holding complex prompts with multiple characters and camera movements, a prompt enhancer for expanding brief descriptions into richer cinematic instructions, and an optional duration predictor that automatically determines clip length from the prompt. You can self-host on your infrastructure with no per-generation billing; commercial use under $10M annual revenue is free under the LTX-2.x Community License. The split model architecture packages each component (transformer, text encoder, VAEs, LoRAs, patches) as individual .safetensors files compatible with ComfyUI and the ltx-pipelines PyTorch package requiring Python >= 3.12, CUDA >= 12.7, and PyTorch ~= 2.7.

Best use cases

Long-form branded video production with consistent character and environment identity. Multishot generation preserves character identity, lighting, and visual style across multiple connected scenes, making this model ideal for corporate videos, product demonstrations, and marketing content that requires multiple shots to feel cohesive without manual stitching or fine-tuning individual clips. The unified audiovisual generation means voice, background sounds, and foley automatically match the visual content’s emotion and pacing, eliminating the need for separate audio post-production workflows.

Audiovisual storytelling for indie creators and animators. The model generates synchronized speech, ambient sound, and foley automatically, removing the burden of sourcing or recording audio separately. Creators can describe complex scenes—”a character running through a forest as the wind picks up and branches snap”—and receive video with matching audio that includes wind, footsteps, and environmental sounds without additional audio engineering.

Product and explainer animation from still images. Image-to-video capability combined with synchronized audio generation enables converting product photography or technical diagrams into animated explainers with narration and sound design in one generation pass. The distilled model’s speed makes iteration practical for creators who need multiple variations.

Robotics and physical AI training data generation. The model is explicitly described as having emerging applicability to robotics and physical AI domains. The world simulation capabilities built into the architecture enable generating diverse, physically plausible action sequences useful for training reinforcement learning models without manual recording.

Multilingual localization of video content. The multilingual text encoder handles prompts in multiple languages, enabling creators to generate region-specific versions of the same story with culturally appropriate audio, accents, and environmental details in a single workflow rather than recording separate audio tracks.

Limitations

Fixed frame-rate and resolution constraints. Frames must satisfy frames % 8 == 1 (valid frame counts: 121, 129, 137, etc.), and width/height must be divisible by 32, restricting flexibility for non-standard aspect ratios or precise duration targeting without awkward padding or cropping.

Multishot consistency imperfect across complex scenes. While multishot generation holds identity across cuts, the model is new in this capability and may still struggle with extreme lighting changes, viewpoint shifts, or complex camera movements that require precise 3D consistency. The diffusion-based approach introduces artifacts in demanding scenes despite the improved decoder.

High memory requirements for full model inference. The full 22B transformer requires substantial VRAM; while the distilled model is faster and lighter, generating long sequences or high-resolution video still demands modern GPUs. CPU offloading via --offload cpu is available but trades speed significantly.

Commercial licensing complexity for organizations above $10M revenue. Organizations with annual revenues exceeding $10M must negotiate a paid Commercial Use Agreement rather than using the Community License, introducing potential friction in larger enterprises. Fine-tune transfers may require paid licensing even under $10M, limiting custom model distribution.

Emerging feature maturity in spatial and audio quality. While significantly improved from LTX-2, the distilled model “retains much more” rather than “all” of the full model’s visual quality, indicating perceptible degradation. Audio quality, though synchronized, is newer than video generation in the Lightricks ecosystem and may not match specialized audio synthesis models like Google’s SoundStorm.

Limited multi-stage upscaling built-in. Spatial and temporal upscalers from LTX-2.3 are still required for multi-stage workflows and not yet packaged in this release, requiring users to chain separate models and manage intermediate outputs manually.

No native support for text overlays or on-screen text in prompts. The model improved text rendering in the decoder, but generating readable on-screen text remains difficult in diffusion-based video generation—users cannot reliably expect captions or UI elements to render legibly.

How it compares

LTX-2 is the immediate predecessor with the same audio-video foundation but without multishot generation or diffusion video decoder improvements. Choose LTX-2.5 for multishot consistency and sharper details; stick with LTX-2 only if already integrated into production and multishot generation is unnecessary. The core tradeoff is that LTX-2.5 is newer with more features but may have less battle-tested edge case handling in production deployments.

LTX-2.3 offers audio-video generation without multishot or diffusion decoder capabilities, trading off scene-to-scene consistency for potentially faster inference if you only need single-shot clips. LTX-2.3 is optimized for quality in individual shots; LTX-2.5 is optimized for connected storytelling. Use LTX-2.3 if your primary need is high-quality single clips with synchronized audio.

ltx-video-0.9.7 generates video only without audio and uses a lighter 13B architecture optimized for 30fps real-time generation. Choose ltx-video-0.9.7 if you need pure speed and have separate audio workflows; choose LTX-2.5 if unified audiovisual generation in a single pass is essential, trading raw inference speed for compositional completeness.

LTX-Video is the full LTX-Video release with video-only focus and real-time 24 FPS generation at 768×512. Use LTX-Video for fastest pure video output when audio is handled separately; use LTX-2.5 when synchronized audio is critical and latency is secondary. LTX-Video is purpose-built for speed; LTX-2.5 is built for audiovisual coherence.

LTX-2.3-fp8 is a quantized version of LTX-2.3 for reduced VRAM footprint, applying to the audio-video problem domain with lower precision. Use LTX-2.3-fp8 if VRAM is severely constrained and single-shot audio-video generation suffices; LTX-2.5 offers multishot at the cost of more memory, but quantized versions of LTX-2.5 are also available (comfy-int8-convrot, NVFP4).

Technical specifications

Architecture. Asymmetric dual-stream transformer (DiT) with 22B-parameter video stream and separate audio stream coupled through bidirectional audio-video cross-attention layers, temporal positional embeddings, and cross-modality AdaLN for shared timestep conditioning. The text encoder is a custom Gemma 4 12B model with projection layers.

Model weights and files. Shipped as split Comfy-aligned pack with separate safetensors files:

Distilled DiT transformer (bf16, 22B parameters): ltx-2.5-22b-distilled-transformer-bf16.safetensors

Full/trainable DiT (bf16): ltx-2.5-22b-dev-transformer-bf16.safetensors

Gemma4 text encoder with projections (bf16): gemma4-12b-with-proj-ltx-2.5-bf16.safetensors

Video VAE (diffusion-based for quality): ltx-2.5-video-vae-bf16.safetensors

Video VAE (conv-based for speed): ltx-2.5-video-vae-conv-bf16.safetensors

Audio VAE with vocoder: ltx-2.5-audio-vae-bf16.safetensors

Duration head patch for auto-length: ltx-2.5-duration-head-bf16.safetensors

Distilled LoRA (450 params): ltx-2.5-22b-distilled-lora-450-bf16.safetensors

Quantization options. bf16 (full precision for PyTorch), fp8-cast (downcast on-the-fly), int8 + convrot (ComfyUI only), and NVFP4 (requires Blackwell GPU or ltx-kernels with ltx-pipelines).

Supported frameworks. ltx-pipelines (PyTorch), ComfyUI (via Comfy int8/NVFP4 checkpoints), and Python API with custom model loading.

Input constraints. Frame counts must satisfy frames % 8 == 1. Width and height must be divisible by 32. Duration can be set explicitly via --num-frames or auto-predicted from prompt via duration head.

Text encoding. Custom Gemma 4 12B with projections handles complex prompts (multiple characters, camera moves, lighting, actions) instead of dropping details. Optional prompt enhancer expands brief prompts into richer cinematic instructions.

Inference optimizations. CPU offloading via --offload cpu trades speed for lower peak VRAM. Fixed 8-step schedule on distilled model with CFG=1.

Output format. MP4 video files with embedded synchronized audio (generated, not externally sourced).

Key new features in 2.5: Native multishot generation, diffusion fidelity rendering with dynamic per-scene compute allocation, diffusion video decoder (replacing VAE), Gemma 4 text encoder, prompt enhancer, optional duration predictor, substantially improved distilled model.

Model inputs and outputs

Inputs

Text prompt: Natural language description of desired video (required); supports complex descriptions with multiple characters, camera movements, lighting, and environmental details.

Image (optional): Frame conditioning for image-to-video; supply via --image PATH FRAME_IDX STRENGTH where FRAME_IDX is the frame position (0 = first) and STRENGTH controls conditioning influence (0.0 to 1.0).

Frame count (optional): --num-frames parameter must satisfy frames % 8 == 1 (e.g., 121, 129, 137). Omit to use optional duration head predictor.

Resolution: Width and height divisible by 32; defaults depend on aspect ratio selection.

Seed: Integer for reproducibility.

Quantization (optional): --quantization fp8-cast for lower VRAM or --quantization nvfp4-prequant for Blackwell GPUs.

Outputs

MP4 file: Video with synchronized generated audio embedded; resolution and duration determined by input parameters or predictor.

Spatial upsampling: Output passes through spatial upsampler (LTX-2.3 checkpoint, currently required as separate model) for final quality.

Getting started

Install and download:

git clone https://github.com/Lightricks/LTX-2.git
cd LTX-2
uv sync
source .venv/bin/activate

hf auth login

# Download LTX-2.5 distilled split pack
hf download Lightricks/LTX-2.5 \
  diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \
  text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
  vae/ltx-2.5-video-vae-bf16.safetensors \
  vae/ltx-2.5-audio-vae-bf16.safetensors \
  model_patches/ltx-2.5-duration-head-bf16.safetensors \
  --local-dir models/ltx-2.5

# Spatial upscaler (required, hosted on LTX-2.3)
hf download Lightricks/LTX-2.3 \
  ltx-2.3-spatial-upscaler-x2-1.1.safetensors \
  --local-dir models/ltx-2.3

Text-to-video generation:

uv run python -m ltx_pipelines.distilled \
  --transformer-path     models/ltx-2.5/diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \
  --text-encoder-path    models/ltx-2.5/text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
  --video-vae-path       models/ltx-2.5/vae/ltx-2.5-video-vae-bf16.safetensors \
  --audio-vae-path       models/ltx-2.5/vae/ltx-2.5-audio-vae-bf16.safetensors \
  --duration-head-path   models/ltx-2.5/model_patches/ltx-2.5-duration-head-bf16.safetensors \
  --spatial-upsampler-path models/ltx-2.3/ltx-2.3-spatial-upscaler-x2-1.1.safetensors \
  --prompt "A golden retriever running through a sunny meadow, cinematic lighting" \
  --seed 42 \
  --output-path output.mp4

Image-to-video with prompt:

uv run python -m ltx_pipelines.distilled \
  --transformer-path     models/ltx-2.5/diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \
  --text-encoder-path    models/ltx-2.5/text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
  --video-vae-path       models/ltx-2.5/vae/ltx-2.5-video-vae-bf16.safetensors \
  --audio-vae-path       models/ltx-2.5/vae/ltx-2.5-audio-vae-bf16.safetensors \
  --duration-head-path   models/ltx-2.5/model_patches/ltx-2.5-duration-head-bf16.safetensors \
  --spatial-upsampler-path models/ltx-2.3/ltx-2.3-spatial-upscaler-x2-1.1.safetensors \
  --image path/to/first_frame.jpg 0 1.0 \
  --prompt "The camera slowly dollies out as wind moves through the grass" \
  --seed 42 \
  --output-path output_i2v.mp4

Python API:

from ltx_pipelines.distilled import DistilledPipeline
from ltx_pipelines.utils.model_paths import ModelPaths

model_paths = ModelPaths.from_split(
    transformer_path="models/ltx-2.5/diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors",
    text_encoder_path="models/ltx-2.5/text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors",
    video_vae_path="models/ltx-2.5/vae/ltx-2.5-video-vae-bf16.safetensors",
    audio_vae_path="models/ltx-2.5/vae/ltx-2.5-audio-vae-bf16.safetensors",
)

pipeline = DistilledPipeline(model_paths)
output = pipeline.generate(
    prompt="Your prompt here",
    num_frames=121,
    seed=42
)

Frequently asked questions

Q: Can I use LTX-2.5 commercially without restrictions?

A: Free commercial use is permitted under the LTX-2.x Community License if your organization has under $10M annual revenue (measured across the whole entity including subsidiaries). Organizations with over $10M revenue must negotiate a paid Commercial Use Agreement. Fine-tune transfers may require a paid license even under $10M.

Q: What GPU VRAM do I need to run LTX-2.5?

A: The full model is memory-intensive. Use the distilled version with fp8-cast quantization and CPU offloading (--quantization fp8-cast --offload cpu) for lower VRAM footprint. Exact VRAM requirements are not specified in the documentation; test on your hardware or use the online playground first.

Q: How does LTX-2.5 differ from LTX-2 for multishot video generation?

A: LTX-2.5 adds native multishot generation in a single pass, generating connected scenes that maintain character identity, environment, lighting, voice, and visual style across cuts. LTX-2 produced single continuous shots only; multishot required post-hoc stitching or separate generation with careful prompt tuning.

Q: Can I fine-tune LTX-2.5?

A: Yes. The full model checkpoint ltx-2.5-22b-dev-transformer-bf16.safetensors is designed for training and fine-tuning in bf16. A distilled LoRA (450 params) is also available. Transfer of fine-tunes may require a paid license depending on your organization’s revenue bracket.

Q: What input formats does the model accept for image-to-video conditioning?

A: The --image flag accepts image file paths (format not specified but standard formats like JPG and PNG should work). You specify the frame index to condition (0 for first frame) and a strength value between 0.0 and 1.0 controlling how strongly the image constrains the generation.

Q: Is LTX-2.5 actively maintained?

A: Yes. LTX-2.5 is the latest release from Lightricks with recent improvements including diffusion fidelity rendering, a new diffusion video decoder, and the duration predictor. The codebase is publicly available on GitHub at https://github.com/Lightricks/LTX-2.

Q: How does frame count work? What values are valid?

A: Frame count must satisfy the constraint frames % 8 == 1. Valid examples: 121, 129, 137, 145 frames, etc. You can omit --num-frames entirely to use the optional duration head predictor, which automatically selects length from the prompt instead of requiring manual specification.

Q: What’s the difference between the diffusion video decoder and the old VAE?

A: The new diffusion video decoder replaces VAE reconstruction, producing sharper faces and textures, better on-screen text rendering, improved motion, and fewer artifacts in demanding scenes. It is the default in LTX-2.5 and represents a quality upgrade over LTX-2.

This is a simplified guide to an AI model called LTX-2.5 maintained by Lightricks. If you like these kinds of analysis, join AIModels.fyi or follow us on Twitter.

Source link