LTX-2.3 22B

'The 22B update to LTX-2: the same 48-block joint audio-video DiT with gated attention, much larger 8-layer text connectors and a bandwidth-extension vocoder for 48 kHz audio. Ships as one bf16 file; ai-toolkit pairs it with a separate Gemma 3 12B text encoder.'

weights: Lightricks/LTX-2.3 ↗
org: Lightricks
modality: video
tasks: text-to-video · image-to-video · text-to-audio-video · image-to-audio-video
license: LTX-2 Community License
released: 2026-03-04not verified
total params: 35.26B
model.arch: ltx2.3

Components

role	model	params	size	dtype	trained
Transformer	LTX-2.3 audio-video DiT huggingface.co/Lightricks/LTX-2.3/blob/main/ltx-2.3-22b-dev.safetensors diffusers.LTX2VideoTransformer3DModel Bundled in ltx-2.3-22b-dev.safetensors (23.07B params, 46.15 GB in total) with every part below except the text encoder. Counted from the model.diffusion_model.* tensors that are not connector weights; size is tensor bytes.	18.99B 18,988,838,144	37.99 GB	bf16+fp32	yes
Text connectors	LTX-2.3 text connectors huggingface.co/Lightricks/LTX-2.3/blob/main/ltx-2.3-22b-dev.safetensors diffusers LTX2TextConnectors The video and audio connector blocks (2.02B) plus the top-level text_embedding_projection (1.16B), which ai-toolkit folds into the connectors. Bundled in the single file.	3.17B 3,172,227,584	6.34 GB	bf16	no
Text encoder	Gemma 3 12B IT (QAT, unquantized) Lightricks/gemma-3-12b-it-qat-q4_0-unquantized transformers.Gemma3ForConditionalGeneration Not in the LTX-2.3 repo. ai-toolkit loads this repo unless te_name_or_path is set. Includes the SigLIP vision tower, which ai-toolkit drops.	12.19B 12,187,325,040	24.37 GB	bf16	no
Tokenizer	Gemma 3 SentencePiece (262k vocab) Lightricks/gemma-3-12b-it-qat-q4_0-unquantized transformers.GemmaTokenizerFast	—	—	—	—
Video VAE	LTX-2.3 video VAE huggingface.co/Lightricks/LTX-2.3/blob/main/ltx-2.3-22b-dev.safetensors diffusers.AutoencoderKLLTX2Video The vae.* tensors of the single file. Smaller decoder than the LTX-2 VAE, same latent space.	726.1M 726,106,225	1.45 GB	bf16	no
Audio VAE	LTX-2 audio VAE huggingface.co/Lightricks/LTX-2.3/blob/main/ltx-2.3-22b-dev.safetensors diffusers.AutoencoderKLLTX2Audio The audio_vae.* tensors of the single file. Same config and parameter count as LTX-2’s.	53.2M 53,248,402	106 MB	bf16	no
Vocoder	LTX-2.3 vocoder with BWE huggingface.co/Lightricks/LTX-2.3/blob/main/ltx-2.3-22b-dev.safetensors diffusers LTX2VocoderWithBWE The vocoder.* tensors of the single file. A 16 kHz vocoder followed by a bandwidth-extension stage that outputs 48 kHz stereo. Only used for sampling.	129.1M 129,085,032	258 MB	bf16	no
total		35.26B	70.52 GB

Latent space

spatial

32×

temporal

8×

channels

128

patch

1×1×1

autoencoder: LTX-2.3 video VAE
pixels per token: 32×32 × 8 frames
frame count: 8n + 1
notes: Same latent space as LTX-2: a 4×4 pixel patchify, then 8× spatial and 8× temporal downsampling. The transformer patch is 1×1×1, so each latent voxel is one token. Width and height must be multiples of 32.

input	latent (c×t×h×w)	tokens
768×768 × 121f	128×16×24×24	9,216	ai-toolkit sample default
768×512 × 121f	128×16×16×24	6,144
1024×1024	128×1×32×32	1,024	still image

Audio latent space

sample rate

16 kHz

temporal

640×

channels

latent rate

25/s

autoencoder: LTX-2 audio VAE
audio channels: stereo
patch: 1 latent steps per token
tokens per second: 25.00
notes: Audio is resampled to 16 kHz and turned into a 64-bin log-mel spectrogram (hop 160, 100 frames/s). The VAE compresses time 4× and mel bins 4×, so each latent step is 8 channels × 16 mel bins, packed into one 128-d token. The BWE vocoder decodes to 48 kHz stereo.

Architecture

Blocks: 48, each with a video and an audio stream
Video stream: 4096 (32 heads × 128), FFN 16384
Audio stream: 2048 (32 heads × 64), FFN 8192
Cross-modal: Audio-to-video and video-to-audio cross-attention in every block, modulated by the other stream’s timestep
Text conditioning: Gemma 3 12B, all 49 hidden states (3840-d) stacked, projected per modality, through 8-layer video and audio connectors, then cross-attention. 1024 tokens, left-padded
Image conditioning: None in the weights. I2V puts the clean first-frame latent in latent frame 0
Timesteps: Per token for video; one timestep for audio
Objective: Rectified flow, resolution-dependent exponential shift (0.95 to 2.05, terminal 0.1)
Changes from LTX-2: Gated attention, cross-attention AdaLN, perturbed-attention support for STG guidance, bigger connectors with per-modality projections, BWE vocoder
Norm / position: QK RMSNorm across heads, split RoPE (3D video with fps, 1D audio)

In AI Toolkit

model.arch: ltx2.3
UI label: LTX-2.3 (video)
model.name_or_path: Lightricks/LTX-2.3/ltx-2.3-22b-dev.safetensors
source: extensions_built_in/diffusion_models/ltx2/ltx2.py
extensions_built_in/diffusion_models/ltx2/convert_ltx2_to_diffusers.py
toolkit/models/v2/diffusion_models/ltx2.py
toolkit/models/v2/vae/ltx2.py
toolkit/models/v2/text_encoders/gemma3.py
extensions_built_in/diffusion_models/ui.tsx
extra UI sections: sample.ctrl_img, datasets.num_frames, model.layer_offloading, model.low_vram, datasets.do_audio, datasets.audio_normalize, datasets.audio_preserve_pitch, datasets.do_i2v, train.audio_loss_multiplier, datasets.auto_frame_count

UI defaults

quantize / quantize_te: true / true
low_vram: true
noise_scheduler / sampler: flowmatch / flowmatch
timestep_type: weighted
sample size: 768×768, 121 frames, 24 fps
train.audio_loss_multiplier: 1.0
datasets.do_audio: true
datasets.do_i2v: false
datasets.cache_latents_to_disk: true
datasets.fps: 24
datasets.auto_frame_count: false
network.conv: disabled (linear LoRA only)

Specifics

Resolution: Buckets and sample sizes snap to multiples of 32.
Frame count: Sampling rounds num_frames down to 8n + 1.
Loading: name_or_path is a single .safetensors file. A repo/file path is looked for in the models folder first and downloaded there if missing. Transformer, connectors, both VAEs and the vocoder are converted from it to Diffusers modules; fp8 files are dequantized first.
Text encoder source: Lightricks/gemma-3-12b-it-qat-q4_0-unquantized, unless te_name_or_path is set (a Gemma 3 folder, repo, or ComfyUI Gemma 3 .safetensors). The vision tower is dropped. Never trained.
Audio: With do_audio on, clip audio is made stereo, resampled to 16 kHz, turned into log-mel and encoded by the audio VAE. The audio stream is noised at the same timestep as video and adds its own loss, scaled by audio_loss_multiplier. Image batches and datasets without audio carry an empty audio stream and no audio loss.
Image-to-video: With do_i2v on, the first frame is encoded and placed in latent frame 0 at t = 0 and masked out of the loss (the mask is renormalized).
Cross timestep: Training and sampling pass use_cross_timestep, the 2.3 behaviour, so each stream’s cross-attention is modulated by the other stream’s sigma.
LoRA target: LTX2VideoTransformer3DModel
Saving: LoRAs are converted to the original LTX-2.3 key layout with the diffusion_model prefix, which ComfyUI can load. Full fine-tunes save the transformer as a Diffusers folder.
Sampling: Samples are mp4 with audio. Adds STG (scale 1, block 28), modality guidance 3, guidance rescale 0.7 and audio guidance 7 on top of guidance_scale. low_vram turns on frame-wise VAE encode and decode.
Metadata base version: ltx2

Example config

not verified

job: extensionconfig:  name: "my_ltx23_lora_v1"  process:    - type: "sd_trainer"      training_folder: "output"      device: cuda:0      network:        type: "lora"        linear: 32        linear_alpha: 32      save:        dtype: float16        save_every: 250        max_step_saves_to_keep: 4      datasets:        - folder_path: "/path/to/videos"          caption_ext: "txt"          caption_dropout_rate: 0.05          num_frames: 121          fps: 24          do_audio: true          do_i2v: false          cache_latents_to_disk: true          resolution: [512, 768]      train:        batch_size: 1        steps: 2000        gradient_checkpointing: true        noise_scheduler: "flowmatch"        timestep_type: "weighted"        audio_loss_multiplier: 1.0        optimizer: "adamw8bit"        lr: 1e-4        dtype: bf16        cache_text_embeddings: true      model:        name_or_path: "Lightricks/LTX-2.3/ltx-2.3-22b-dev.safetensors"        arch: "ltx2.3"        quantize: true        quantize_te: true        low_vram: true      sample:        sampler: "flowmatch"        sample_every: 250        width: 768        height: 768        num_frames: 121        fps: 24        guidance_scale: 3        sample_steps: 25        prompts:          - "a bear building a log cabin in the snow covered mountains, the sound of an axe chopping wood"