⬡ LTX Video Generator¶

Generates LTX-2.3 video+audio using a dual-pass pipeline with ManualSigmas.

Unified Node

As of v2.5, ⬡ LTX Video Generator is an alias to the unified ⬡ Video Generator node. Both WAN and LTX models are supported by the same node — the pipeline is auto-selected based on the model bundle's loader_type. Existing workflows using the LTX Video Generator node will continue to work unchanged.

Inputs¶

Name	Type	Required	Description
`model_bundle`	`UME_BUNDLE`	✅	From LTX Loader or Bundle Auto-Loader
`positive`	`POSITIVE`	✅	Describe the scene/action for the video
`video_settings`	`UME_VIDEO_SETTINGS`	✅	From LTX Video Settings node
`negative`	`NEGATIVE`	❌	Describe what to avoid. LTX-2.3 uses ConditioningZeroOut by default
`loras`	`UME_LORA_STACK`	❌	Connect a LoRA Block node
`source_image`	`IMAGE`	❌	Source image for Image-to-Video (I2V) mode

Outputs¶

Name	Type	Description
`video_pipe`	`UME_VIDEO_PIPELINE`	Complete video pipeline with frames + optional audio, ready for Video Output

Pipeline¶

The LTX Video Generator orchestrates the entire LTX-2.3 pipeline internally:

graph LR
    A["LoRA Application"] --> B["CLIP Text Encode"]
    B --> C["ConditioningZeroOut (negative)"]
    C --> D["LTXVConditioning (frame_rate)"]
    D --> E["Empty Latents (video+audio)"]
    E --> F["Pass 1: Low-res Sampling"]
    F --> G["Latent Upscale 2x"]
    G --> H["Pass 2: Full-res Sampling"]
    H --> I["Tiled VAE Decode (video)"]
    H --> J["Audio VAE Decode"]

Text-to-Video (T2V)¶

Connect only the positive prompt — no source image needed.

Image-to-Video (I2V)¶

Connect a source image to condition the first frame via LTXVImgToVideoInplace.

Dual-Pass vs Single-Pass¶

With upscaler: Pass 1 at half resolution → 2x upscale → Pass 2 at full resolution (best quality)
Without upscaler: Single pass at full resolution

Audio¶

When audio_enabled=True in video settings and an Audio VAE is loaded: - Empty audio latents are created alongside video latents - Audio+Video latents are combined via NestedTensor for joint sampling - After sampling, audio is decoded separately and muxed into the output video

Connection to Video Output¶

⬡ LTX Loader → ⬡ LTX Video Settings → ⬡ LTX Video Generator → ⬡ Video Output
                                              ↑
                                     ⬡ Positive Prompt
                                     ⬡ LoRA Block (optional)
                                     IMAGE (optional, for I2V)

CFG = 1.0

LTX-2.3 distilled models use cfg=1.0 internally. The node handles this automatically.

VRAM Usage

Video decode uses spatio-temporal tiling (4 spatial tiles + temporal chunking) to keep peak VRAM low even for long, high-resolution videos.