How to Choose the Right LTX-2 Pipeline for Your Specific Use Case

LTX-2 provides multiple specialized pipelines that trade off speed, quality, conditioning options, and HDR support—your choice depends on whether you need video editing, audio-driven generation, HDR output, or maximum fidelity.

Selecting the right LTX-2 pipeline requires matching your project's technical constraints against the library's available generation modes. Lightricks designed each pipeline for distinct scenarios: single-stage speed, two-stage quality, mask-based editing, or specialized conditioning like audio or keyframe interpolation. This guide maps concrete use cases to their optimal pipeline implementations based on the source code in packages/ltx-pipelines/src/ltx_pipelines/.

Editing Existing Video: RetakePipeline

If you need to regenerate a specific time region while preserving the rest of your video, RetakePipeline is purpose-built for this workflow.

Implemented in packages/ltx-pipelines/src/ltx_pipelines/retake.py, this pipeline performs single-stage, mask-based regeneration. It accepts both video and audio inputs, making it suitable for replacing visual content without re-rendering unaffected segments.

uv run python -m ltx_pipelines.retake \
    --checkpoint-path models/ltx-2.5/diffusion_models/ltx-2.5-22b-dev.safetensors \
    --video-path source.mp4 \
    --start-time 5.0 \
    --end-time 8.0 \
    --prompt "Replace the middle scene with a flying dragon" \
    --output-path edited.mp4

Audio-Driven Video Generation: A2VidPipelineTwoStage

For synchronizing video to audio tracks, A2VidPipelineTwoStage provides dedicated audio conditioning. Located in packages/ltx-pipelines/src/ltx_pipelines/a2vid_two_stage.py, this pipeline runs two stages: Stage 1 generates low-resolution video conditioned on encoded audio features, and Stage 2 upsamples to final resolution.

This architecture ensures temporal coherence with the audio waveform rather than treating sound as an afterthought.

uv run python -m ltx_pipelines.a2vid_two_stage \
    --checkpoint-path models/ltx-2.5/diffusion_models/ltx-2.5-22b-dev.safetensors \
    --distilled-lora models/ltx-2.5/loras/ltx-2.5-22b-distilled-lora.safetensors \
    --spatial-upsampler-path models/ltx-2.5/latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2.safetensors \
    --audio-path mytrack.wav \
    --prompt "A serene forest animated to the music" \
    --output-path audio_to_video.mp4

HDR and Professional Output Workflows

LTX-2 supports multiple HDR pipelines depending on your color space requirements. The documentation in packages/ltx-pipelines/docs/hdr.md specifies native --hdr support with options for SRGB_LINEAR, ACESCG, and ACESCCT.

Standard HDR Pipelines

If you have EXR plates or need native HDR output, use DistilledPipeline, TI2V, RetakePipeline, or ICLoraPipeline with the --hdr flag:

--hdr SRGB_LINEAR  # or ACESCG, ACESCCT

Advanced HDR with Custom Tonemapping

For workflows requiring linear HDR float data for custom color grading, HDRICLoraPipeline in packages/ltx-pipelines/src/ltx_pipelines/hdr_ic_lora.py outputs linear EXR files directly. This bypasses built-in tonemapping for compositing pipelines.

Video-to-Video and Keyframe Workflows

Two pipelines handle reference-based generation with distinct input modalities.

Reference Video Transformations: ICLoraPipeline

ICLoraPipeline (packages/ltx_pipelines/src/ltx_pipelines/ic_lora.py) applies video-to-video transformations using IC-LoRA conditioning. Provide a source video and prompt to restyle or modify content while maintaining structural consistency.

Image Keyframe Interpolation: KeyframeInterpolationPipeline

For generating video between static keyframe images, KeyframeInterpolationPipeline in packages/ltx-pipelines/src/ltx_pipelines/keyframe_interpolation.py accepts multiple reference frames and interpolates smooth motion between them.

Speed vs. Quality Trade-offs

Maximum Fidelity: TI2VidTwoStagesPipeline

When production quality is paramount and inference time is secondary, TI2VidTwoStagesPipeline (packages/ltx_pipelines/src/ltx_pipelines/ti2vid_two_stages.py) provides multimodal guidance with dedicated upsampling. The HQ variant, TI2VidTwoStagesHQPipeline, adds further refinement layers.

uv run python -m ltx_pipelines.ti2vid_two_stages_hq \
    --checkpoint-path models/ltx-2.5/diffusion_models/ltx-2.5-22b-dev.safetensors \
    --distilled-lora models/ltx-2.5/loras/ltx-2.5-22b-distilled-lora.safetensors \
    --spatial-upsampler-path models/ltx-2.5/latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2.safetensors \
    --prompt "A photorealistic underwater city" \
    --output-path hq_output.mp4

Fastest Inference: DistilledPipeline

For batch workloads where speed matters, DistilledPipeline (packages/ltx_pipelines/src/ltx_pipelines/distilled.py) uses a distilled model with a fixed 8-sigma noise schedule. This is the default recommendation for most production deployments requiring throughput over absolute quality.

uv run python -m ltx_pipelines.distilled \
    --checkpoint-path models/ltx-2.5/diffusion_models/ltx-2.5-22b-distilled.safetensors \
    --distilled-lora models/ltx-2.5/loras/ltx-2.5-22b-distilled-lora.safetensors \
    --spatial-upsampler-path models/ltx-2.5/latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2.safetensors \
    --prompt "A sunrise over a futuristic cityscape" \
    --output-path output.mp4

Detail Enhancement and Temporal Upsampling

DFRPipeline in packages/ltx_pipelines/src/ltx_pipelines/dfr_pipeline.py runs a Diffusion Fidelity Rendering LoRA for maximum detail recovery. It optionally performs temporal upsampling at 2× or 4× frame rates, useful when the base model's framerate is insufficient for your delivery requirements.

Specialized Single-Purpose Pipelines

Rapid Prototyping: TI2VidOneStagePipeline

TI2VidOneStagePipeline (packages/ltx_pipelines/src/ltx_pipelines/ti2vid_one_stage.py) skips the upsampling stage for fastest iteration during development. Quality is reduced but generation completes in roughly half the time of two-stage variants.

Audio-Only Generation: T2AOneStagePipeline

When you need text-to-audio without video, T2AOneStagePipeline (packages/ltx_pipelines/src/ltx_pipelines/t2a_one_stage.py) runs the audio generation branch in isolation.

Lip-Sync Dubbing: DubItPipeline

For replacing dialogue while preserving mouth movements, DubItPipeline (packages/ltx_pipelines/src/ltx_pipelines/dubit.py) applies Dub-It IC-LoRA conditioning across both generation stages.

Pipeline Selection Decision Framework

The definitive decision tree lives in packages/ltx-pipelines/docs/pipeline-selection.md. This document codifies the mapping between technical requirements and pipeline selection.

For complete API reference including stage configurations, conditioning parameters, and model compatibility, consult packages/ltx-pipelines/docs/pipelines.md.

Summary

• RetakePipeline — Edit existing video segments without full regeneration
• A2VidPipelineTwoStage — Synchronize video generation to audio input
• HDRICLoraPipeline — Linear EXR output for professional color workflows
• ICLoraPipeline — Video-to-video transformation with reference conditioning
• KeyframeInterpolationPipeline — Animate between static keyframe images
• TI2VidTwoStagesPipeline — Maximum quality for production delivery
• DistilledPipeline — Fastest inference for batch and real-time workflows
• DFRPipeline — Enhanced detail with optional temporal upsampling
• TI2VidOneStagePipeline — Rapid prototyping without quality upsampling
• T2AOneStagePipeline — Audio-only generation when video is not required
• DubItPipeline — Dialogue replacement with preserved lip synchronization

Frequently Asked Questions

What is the fastest LTX-2 pipeline for batch processing?

DistilledPipeline provides the fastest inference by using a distilled model with a fixed 8-sigma schedule. According to the LTX-2 source code, this pipeline is explicitly recommended as the default for most batch workloads where throughput outweighs absolute visual fidelity.

How do I generate HDR video with LTX-2?

For standard HDR output, use DistilledPipeline, TI2V, RetakePipeline, or ICLoraPipeline with the --hdr flag set to SRGB_LINEAR, ACESCG, or ACESCCT. For linear EXR float data requiring custom tonemapping, use HDRICLoraPipeline which outputs directly to linear color space without built-in curve application.

Can LTX-2 edit a specific portion of an existing video?

Yes. RetakePipeline in packages/ltx_pipelines/src/ltx_pipelines/retake.py performs mask-based regeneration on specified time regions (via --start-time and --end-time) while preserving unselected video segments. It handles both video and audio channels in the edited region.

Which pipeline should I choose for audio-reactive video generation?

Use A2VidPipelineTwoStage. This pipeline encodes audio features in Stage 1 to condition low-resolution generation, then upsamples in Stage 2 with audio-guided temporal coherence. The conditioning is integrated rather than post-processed, ensuring genuine audio-visual synchronization.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}