Explore LTX-2.3

Provides the exact transcript the speech model will read aloud. Write the words, punctuation, and language you want spoken; punctuation can influence pauses and phrasing, while pronunciation still depends on the model’s training data.

Describes unwanted subjects, styles, defects, or artifacts such as blur or watermarks. It steers generation away from those concepts but is not a guaranteed content filter.

Provides instructions to the optional prompt-enhancement model that rewrites the user prompt before generation. Supplying it enables enhancement; use it to define the desired expansion style, or leave it empty when the original wording must remain unchanged.

Sets the horizontal resolution of generated images or video frames in pixels. Larger values create more detail but require more memory and processing time; combine it with height to choose the aspect ratio.

Sets the vertical resolution of generated images or video frames in pixels. Larger values create more detail but require more memory and processing time; keep the height compatible with the model’s supported dimensions and desired aspect ratio.

Sets the playback rate used for the generated video and synchronized audio. Together with the frame count it determines duration—for example, 120 frames at 24 fps is about five seconds—and higher rates require more frames for the same length.

Controls the length of the generated video as a frame count. Duration depends on the playback frame rate, and more frames increase generation time and memory use; some pipelines accept only particular frame-count intervals.

Controls how many iterative denoising passes produce the final media. More steps can refine detail and consistency but increase generation time, with diminishing quality gains beyond the model’s useful range.

Overrides the scheduler’s default noise-level sequence with custom sigma values. This changes how denoising effort is distributed and can strongly affect quality; use values designed for the selected scheduler and do not combine them with custom timesteps.

Overrides the scheduler’s normal diffusion timestep sequence with a custom descending list. This is an advanced reproducibility and scheduler-tuning control; use a schedule supported by the selected scheduler and do not combine it with custom sigmas.

Sets the strength of classifier-free guidance toward the text prompt. Higher values usually improve literal prompt adherence but can reduce natural motion or create artifacts; lower values give the model more visual freedom.

Controls spatio-temporal guidance for video consistency across space and time. Higher values can strengthen coherent motion and structure but may reduce variation or introduce artifacts; zero disables this guidance path.

Controls how strongly the video generation branch follows cross-modal conditioning such as the prompt and synchronized audio context. Higher values increase conditioning influence, while excessive values can reduce natural motion or visual quality.

Rescales classifier-free guidance to reduce overexposure, excessive contrast, or color saturation at higher guidance strengths. Zero disables rescaling, while larger values apply more correction and may soften prompt adherence.

Sets classifier-free guidance strength for generated audio. Higher values make dialogue, effects, and ambience follow the prompt more literally, but excessive guidance can reduce naturalness or introduce audible artifacts.

Controls spatio-temporal guidance in the audio generation branch to encourage consistency over time. Higher values may improve continuity but can constrain variation; zero disables this additional guidance.

Controls how strongly generated audio follows cross-modal video and prompt conditioning. Increase it when sound is weakly connected to the scene, but excessive values may create unnatural or overemphasized audio.

Rescales audio classifier-free guidance to limit distortion caused by strong guidance. Zero disables correction, while larger values temper the guided prediction and can trade strict prompt adherence for cleaner sound.

Lists the transformer block indices where spatio-temporal guidance is applied. Block 28 is recommended for this checkpoint; changing the list alters which representation layers receive guidance and is intended for advanced model tuning.

Sets how much initial noise is added before synchronized media generation. Higher values introduce more randomness and variation, while zero relies on the standard initialization; changing it can affect reproducibility and scene stability.

Requests multiple video variations from one prompt in a single run. Each result uses additional computation and memory, and different seeds or random states allow the variations to differ.

Selects the normalized diffusion timestep used when decoding generated latents into media. The setting affects the balance between denoising and retained variation and should normally remain at the pipeline default unless reproducing a tested workflow.

Controls additional noise applied during latent decoding. Leave it automatic for the pipeline’s recommended behavior; explicit values can change texture and detail but may also introduce instability or reduce reproducibility.

Allows guidance information to be shared across diffusion timesteps instead of treating every step independently. This can improve temporal consistency in synchronized media, while disabling it provides the simpler per-step behavior.

Limits how many tokens the optional prompt-enhancement model may add while rewriting the request. A larger budget allows richer scene and audio detail but increases preprocessing time and may drift farther from the original prompt.

Passes JSON generation options such as temperature directly to the prompt-enhancement model. These settings affect the rewritten prompt rather than final media diffusion, and unsupported keys can cause preprocessing errors.

Seeds the optional prompt-enhancement model for repeatable prompt rewrites. Reuse it to obtain the same expansion under identical settings, or change it to explore a different interpretation before media generation.

Caps the tokenized prompt length accepted by the model’s text encoder. Text beyond the limit is truncated, so increasing it can preserve long descriptions when supported but uses more memory.

Passes advanced JSON options directly to the model’s attention processor. Use only arguments supported by the installed pipeline, because these values can alter attention scaling, performance, or memory use and invalid keys may fail inference.

Initializes the pipeline’s random-number generator. Reusing the same seed, prompt, model, and settings usually reproduces the same output, while changing it creates a new variation.

Chooses the data representation returned by the pipeline. Image frames are convenient for viewing, NumPy arrays suit numerical processing, and latent tensors are compact intermediate data that still require decoding.

Chooses the structure of the pipeline result. Enabled returns an object with named properties that is easier to read; disabled returns a positional tuple that may suit older or lower-level integrations.