LTXVideo

LTXVideo is a family of open-weight video models from Lightricks. From LTX-2.0 onward, picture and synchronized audio are generated jointly from a single prompt, so every clip carries a soundtrack whether or not one was described.

SD.Next drives LTX through a dedicated Video -> LTXVideo tab which adds a two-stage upsample and refine workflow, per-variant scheduler and guidance correctness, pruned downloads, and pre-quantized model rows.

Note

Every supported LTX model is listed in the LTXVideo tab. LTX is deliberately excluded from the generic Video -> Models tab, so the LTXVideo tab is the only place these models are loaded from.

Important

Video support requires ffmpeg to be installed and available in the PATH

Important

Models are downloaded automatically on first use and cached. Manually downloaded LTX models are not supported.

Model Families

Four generations are supported. Newer is not strictly a superset: auto duration exists only on LTX-2.5, and the per-frame decode controls exist only on LTX-0.9.x.

Family	Model	Text encoder	Audio	Notes
LTX-2.5	22B	Gemma 4	48 kHz	Current generation, auto duration, gated download
LTX-2.3	22B	Gemma 3	48 kHz	v1.0 and v1.1 rows, pre-quantized rows available
LTX-2.0	19B	Gemma 3	24 kHz	First generation with joint audio and video
LTX-0.9.x	2B or 13B	T5-XXL	none	Legacy, video only

Each family is listed under a separator in the model dropdown. Selecting a separator does nothing: pick an actual model below it.

LTX-2.5

LTX-2.5 is the current generation and the recommended starting point. The transformer is a 19B DiT paired with roughly 3B of text connectors, and the text encoder is a fine-tuned Gemma 4 that is version-locked to the checkpoint. A stock Gemma 4 is not accepted in its place.

What LTX-2.5 adds over LTX-2.3:

Duration head: clip length can be predicted from the prompt instead of set by hand
Native multi-shot: several shots with explicit cuts described in one prompt
Higher image conditioning fidelity: input images are re-compressed at CRF 18 rather than 33
In-repo spatial upsampler and refine LoRA, so the two-stage workflow needs no third-party mirror

Six rows are available, covering text-to-video, image-to-video and conditioning in both Distilled and Dev variants.

Important

The LTX-2.5 weights are gated. Accept the license at Lightricks/LTX-2.5-Diffusers and configure an access token before selecting one of these rows. See Gated Models for the procedure.

LTX-2.3

LTX-2.3 remains useful as the host of every released IC-LoRA checkpoint, and it is the only current-generation family with pre-quantized rows. Both v1.0 and v1.1 are listed, and each has SDNQ pre-quantized rows.

LTX-2.0

The original joint audio and video generation. Prompt grammar is unchanged from later releases; adherence, audio quality and detail are lower. Audio is produced at 24 kHz here, against 48 kHz on 2.3 and 2.5.

LTX-0.9.x

Legacy releases from 0.9.0 through 0.9.8, ranging from a 2B model to a 13B one. These produce video only, use a T5-XXL text encoder, and are the only family where decode timestep and image cond apply.

Downloads and Storage

Only the components a given row actually calls are fetched. The transformer is loaded separately and its folder is pruned from the pipeline download, so a Dev row and a Distilled row of the same family cost the same.

Approximate first-use download for one LTX-2.5 row is 72GB, made up of the transformer at 38GB, the text encoder at 24GB, the connectors at 6.3GB, and roughly 3GB of VAE, audio VAE, vocoder, spatial upsampler and duration head.

Selecting a second row from the same family reuses everything except the transformer.

Tip

Enable Refine on a Dev row and an additional 9.7GB distilled refine LoRA is fetched on first use.

Hardware

LTX-2.x models are large. A GPU with at least 16GB VRAM and a system with 64GB RAM is the practical floor, and the two-stage workflow at high resolution asks for more.

Quantization

SDNQ pre-quantized rows are listed for LTX-2.0 and LTX-2.3 and are marked SDNQ-4Bit in the dropdown. These cut the download roughly in half and load directly in uint4 with no quantization pass at load time.

Where no pre-quantized row exists, quantization-on-the-fly can be enabled instead.

See Quantization and SDNQ Quantization for details

Offloading

Balanced offloading is the SD.Next default and is a reasonable starting point.

For LTX-2.x at high resolution, group offloading with leaf level and record stream both enabled is what keeps a 24GB card working. Block-level group offload thrashes on the same workload.

See Offloading for details

Dev and Distilled

Every LTX-2.x family ships two transformer variants, and the choice changes how the sampler is driven.

	Dev	Distilled
Denoise	30 guided steps	8 fixed sigmas, unguided
Guidance scale	3.0	1.0
Dynamic shift	on	off
Extra guidance terms	four-way composition	none, baked into the schedule

Selecting a model applies the matching defaults automatically.

Note

The steps slider has no effect on a Distilled row. Distilled inference is driven by an explicit sigma schedule rather than a step count, and handing it a step count would substitute a generic linear schedule and cost quality.

Tip

Compose on Distilled and re-render the keeper on Dev.

Workflows

The workflow is chosen by the model row, and the UI shows only the inputs that row accepts.

T2V generates from the prompt alone.

I2V requires an input image, supplied under Input media. The image anchors the opening frame.

Condition accepts a richer set of conditioning inputs and is the most flexible row:

Image: anchors the opening frame
Last image: anchors the closing frame, giving a first-to-last-frame workflow
Video prefix: a source video anchors the opening frames, with controls for how many frames to take and how many to skip between them
Gallery prefix: a batch of images anchors the opening frames instead of a video

Input strength sets how strongly each conditioning input binds, from 0.0 to 1.0.

Note

Conditioning images are re-compressed before use, at the H.264 quality the model was trained against: CRF 18 on LTX-2.5, CRF 33 on LTX-2.0 and LTX-2.3. This is intentional and matches the reference implementation.

Parameters

Resolution and Frames

Width and height are rounded down to a multiple of 32. With two-stage refine active the final resolution must divide by 64, and a resolution that does not is adjusted down with a warning in the log.

Frame count is quantized to 8n + 1, so the default 121 frames is 121 and yields roughly 5 seconds at 24 fps.

Tip

The LTX-2.5 reference example runs 960x544 at 121 frames, and the SD.Next default is 1024x576. Resolution and frame count both feed the same token sequence, so cost rises steeply in either direction.

Auto Duration

Available on LTX-2.5 only.

Tick auto duration and the clip length is predicted from the prompt by the duration head instead of read from the frames slider. Predicted lengths are quantized to the VAE temporal grid, so the realized duration lands on roughly 0.33 second boundaries at 24 fps, and are bounded between 1 and 20 seconds.

The realized length is carried into upscale and refine rather than predicted a second time.

Guidance and Steps

Guidance scale and steps are set from the model defaults whenever the model is changed. Leaving them alone is correct for most work.

Several further guidance terms are applied automatically and have no sliders: spatio-temporal guidance and its block selection, modality scale, guidance rescale, the audio guidance stack, and the cross-timestep flag that 2.3 and 2.5 need. These are set per variant, since the correct values differ between families and between Dev and Distilled, and a wrong value produces flicker or striping rather than an obvious failure.

Sampler Shift and Dynamic Shift

Sampler shift at -1 keeps the model default. Any value from 0 upward overrides it.

Dynamic shift is set automatically: on for Dev rows, off for Distilled rows and for LTX-0.9.x.

Warning

Enabling dynamic shift on a Distilled row remaps its fixed sigma schedule and degrades output.

Decode Timestep and Image Cond

These two apply to LTX-0.9.x only and are inactive on LTX-2.x, which decodes through a different path.

Upscale and Refine

Refine is enabled by default on every LTX-2.x row and is the single largest quality lever on the page.

With Refine on and Upscale off, the canonical two-stage recipe runs:

Stage 1 denoises at half the requested resolution
The latents are upsampled 2x by the spatial upsampler
Stage 2 refines at the requested resolution over three fixed sigmas at identity guidance

On Dev rows, stage 2 loads the distilled refine LoRA so the second stage runs the distilled recipe over the dev weights. On Distilled rows the second stage runs natively.

Upscale without Refine upsamples the latents and decodes them, with no second denoise pass. The scale factor is adjustable from 1.0 to 4.0.

Strength applies to LTX-0.9.x conditioning pipelines only. LTX-2.x refine runs fixed canonical schedules and ignores it.

Tip

Ask for the resolution you want as the final output. Stage 1 halves it internally.

Note

The refine stage runs at the full requested resolution, so peak memory is set there rather than by the half-resolution first stage.

Audio

Audio is generated jointly with the picture on every LTX-2.x row and is on by default. Save audio under the Audio accordion controls whether it is muxed into the output file.

Audio is written at the rate the loaded vocoder reports: 48 kHz on LTX-2.3 and LTX-2.5, 24 kHz on LTX-2.0.

Note

Turning Save audio off skips the vocoder pass but does not make the model silent internally. Audio cross-attention still runs during denoising, because the picture is conditioned on it.

Tip

A prompt that describes no sound still produces sound, just sound that was not chosen. Describe the ambience, the music and any dialogue.

LoRA

LTX LoRAs are used the same way as any other network: select one from the Networks tab or add <lora:name:strength> to the prompt.

Warning

LoRAs are generation-specific. An LTX-0.9.x LoRA does not load onto LTX-2.x, and the reverse is also true. Within LTX-2.x, most 2.3 adapters load onto 2.5, but validate one before relying on it.

LTX runs through the Diffusers loading method rather than the native SD.Next loader, so LoRA-related settings under Settings -> Extra Networks -> LoRA that name the native path do not apply here.

Adapters that carry text connector deltas are re-routed to the connectors on load. Without this, those weights land in the transformer namespace and are silently dropped, which shows up as jumbled dialogue rather than as an error.

Official adapters worth knowing:

LTX-2.3 Creative Lab: the released LoRA and IC-LoRA checkpoints, covering relight, colorization, deblur, in-outpainting, cinemagraph and foley
LTX-2 collection: camera control adapters for dolly, pan and orbit moves, plus the base checkpoints
The distilled refine LoRA ships inside the LTX-2.5 model repo and is fetched automatically by the refine stage

Note

IC-LoRAs require the in-context pipeline, which the LTXVideo tab does not currently expose. Style, motion and identity LoRAs load normally.

Tip

On a Dev row with Refine enabled, the refine stage activates its own adapter, so a prompt LoRA shapes stage 1.

Prompting

LTX reads one flowing paragraph in present tense, not a shot list and not JSON. The working formula is Shot + Scene + Action + Character + Camera + Audio.

Numeric direction such as "pans right at 2 degrees per second" is a documented mistake. Describe the move in words
Negative prompts are not part of the recommended approach. Steer positively and stay internally consistent
Spoken lines go in quotation marks, split into short quoted phrases with acting directions between them
On LTX-2.5, prompt length should match scene complexity. On earlier generations it should scale with clip duration
Multi-shot, written as one chronological paragraph with explicit cuts, is LTX-2.5 only. Earlier generations produce a single continuous take

A full prompting guide covering both generations, conditioning inputs and screenplay form is maintained at LTX prompting and usage guide. The same repo publishes a model reference table of native resolutions, durations, frame rates and step counts.

Tip

The Prompt Enhance script under Video -> Extras expands a sparse brief into richer cinematic instruction. Write the prompt out in full when exact wording matters.

Output and Encoding

Encoding, interpolation, frame saving and upscaling are shared across every video model and are configured under Video -> Output.

Video files are written to the folder set in Settings -> Image Paths -> Folder for videos.

See Video for encoding, interpolation and upscaling details

Troubleshooting

Download fails with a 401 or an access error The LTX-2.5 rows are gated. Accept the license and configure a token, see Gated Models.

Output is oversaturated or striped Check that Dynamic shift matches the variant. It should be off on Distilled rows.

Resolution came out smaller than requested Two-stage refine needs a final resolution divisible by 64. The adjustment is logged.

Dialogue is present but the words are scrambled A LoRA carrying connector weights failed to route. Confirm the adapter matches the model generation.

Generation is very slow or the GPU stalls Switch to group offloading with leaf level and record stream enabled, see Offloading.

For any LTX issue, set SD_VIDEO_DEBUG=true, reproduce, and attach the full sdnext.log to a GitHub issue.

License

LTX-2.x is released under the LTX-2.x Community License, which permits commercial and production use at no cost below a revenue threshold. LTX-0.9.x releases each carry their own license file in the respective repository.