LTX 2.3 is a 22-billion-parameter open-source AI video model built by Lightricks on the Diffusion Transformer (DiT) architecture. It supports text-to-video, image-to-video, audio-to-video, and video-to-video generation with native portrait output, rebuilt VAE, and a 4x-expanded text connector for more accurate prompt interpretation. Model weights are available on Hugging Face for both dev and distilled checkpoints.

Do I need a powerful GPU or local desktop setup?

No. On ltx23.app all rendering happens in the cloud — no local GPU, VRAM, or desktop installation required. If you prefer running locally, LTX 2.3 supports ComfyUI workflows and GGUF/FP8 quantized formats for lower hardware requirements. The recommended local setup is an NVIDIA GPU with 32 GB+ VRAM, 32 GB RAM, and 60 GB storage on Windows.

How does LTX 2.3 compare to other video models like WAN 2.2?

On H100 GPUs, the LTX 2 series achieves roughly 18x the throughput of WAN 2.2 14B, making it significantly faster for batch rendering. LTX 2.3 also introduces native 9:16 portrait video, a reworked audio vocoder, and sharper edge detail from its rebuilt latent space — improvements not yet matched by most competing open-source models.

What video specs does the model support?

Videos render at up to 1080p HD in 16:9, 9:16, 1:1, and 4:3 aspect ratios with durations from 4 to 20 seconds including audio-synced output. LTX 2.3 is the first in its line to support native vertical 1080x1920, trained on real portrait data rather than cropped landscape. Prompts accept up to 2,000 characters for detailed scene descriptions.

Is LTX 2.3 free to use?

Yes. New accounts on ltx23.app receive free credits to try AI video generation. After that, you can purchase additional credits or subscribe to a plan to keep creating. Subscription plans offer volume discounts for frequent creators.

Can I use LTX 2.3 outputs commercially?

Yes. Videos generated on ltx23.app include full commercial rights — no watermarks, no royalty fees. The open-source license also permits commercial use of locally generated outputs for qualifying organizations, covering advertising, social media, broadcast, and print.

Which model formats and workflows are available?

LTX 2.3 is available as a base checkpoint, distilled checkpoint with LoRA, FP8 scaled variant, and GGUF quantized format. It integrates directly with ComfyUI for custom workflows including first-and-last-frame control, spatial upscaler, depth conditioning, and IC-LoRA motion tracking. All weights are downloadable from Hugging Face.

How do I get started with LTX 2.3?

Create a free account on ltx23.app, enter a text prompt describing your video, optionally upload a reference image or audio, set parameters like duration and aspect ratio, then click generate. Your video is ready to download in moments — no video editing or AI expertise required.

LTX 2.5 ComfyUI: The Three Native Workflows Explained

ComfyUI support for LTX-2.5 was ready on day one: three native workflows ship in the Template Library — T2V, I2V, and FLF2V — with no custom nodes. If you ran LTX 2.3 in ComfyUI, the good news is the muscle memory transfers. The bad news is every file changed, and one new step will block you if you miss it: the 2.5 weights live in a gated Hugging Face repository, and downloads fail until your access request is approved.

This guide walks the official setup end to end: access, files, folders, the three workflows, and the prompting tips that come with them. Sourced from the ComfyUI LTX-2.5 tutorial, the LTX-2.5 model card, and the LTX-2 README. Checked August 13, 2026.

Step 0: Unlock the gated repo first

The official ComfyUI docs are explicit: "The LTX-2.5 model files are hosted in the gated Lightricks/LTX-2.5 repository… Model downloads will fail without access."

Open the repository page
Accept the license agreement (LTX-2.x Community License)
Wait for your access request to be approved

Access requests are reviewed by Lightricks — the turnaround time is not documented, so request access before you do anything else. This is the step that produces the classic "node is red / model not found" on a fresh setup when the real problem is a 401 at download time. The GitHub README has the same warning for the CLI path: accept the terms and log in with a Read token (fine-grained tokens need the "read gated repos" scope enabled).

Step 1: Update ComfyUI and load a template

Update ComfyUI to the latest version (Nightly has the newest core nodes)
Open the Template Library → Video → choose any LTX-2.5 workflow
Missing nodes usually mean an outdated ComfyUI — update, then reload

The three official workflows:

Workflow	What it does	Extra files vs. T2V
T2V	Text-to-video with optional spatial upscaling	—
I2V	Animates an input first-frame image	Your image (default `neon_cyborg_portrait.png` provided)
FLF2V	Interpolates between a first and last frame	First + last frame images (defaults provided)

Step 2: Download every file to its exact folder

The T2V and I2V workflows use the same six files. FLF2V uses the same set minus the spatial upscaler.

Model file	ComfyUI folder
`ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors`	`models/diffusion_models/`
`gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors`	`models/text_encoders/`
`gemma4_e2b_it_bf16.safetensors` (from Comfy-Org/gemma-4)	`models/text_encoders/`
`ltx-2.5-video-vae-bf16.safetensors`	`models/vae/`
`ltx-2.5-audio-vae-bf16.safetensors`	`models/vae/`
`ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors`	`models/latent_upscale_models/`

Complete storage layout from the official tutorial:

ComfyUI/
└── models/
    ├── diffusion_models/
    │   └── ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors
    ├── text_encoders/
    │   ├── gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors
    │   └── gemma4_e2b_it_bf16.safetensors
    ├── vae/
    │   ├── ltx-2.5-video-vae-bf16.safetensors
    │   └── ltx-2.5-audio-vae-bf16.safetensors
    └── latent_upscale_models/
        └── ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors

Three things that trip people up:

Two text encoders, not one. The LTX-specific gemma4-12b-with-proj-ltx-2.5 encoder is required, but the official workflow also loads a second encoder, gemma4_e2b_it_bf16, from the Comfy-Org repo. Download both.
These are the Comfy int8 builds. The comfy-int8-convrot variants are ComfyUI-only — the model card explicitly says they are not for the ltx-pipelines / PyTorch path. If you also use ltx-pipelines, download the bf16 files for that path instead. For ComfyUI, int8 is the intended default: smaller footprint, and the reason 8 GB-class VRAM runs are being reported in the community within hours of release (preliminary reports — verify on your own card).
Audio VAE is not optional in these workflows. The LTX-2 family generates synchronized audio, and the audio VAE (ltx-2.5-audio-vae-bf16.safetensors) is in the official file list. Skip it and audio nodes go red.

Step 3: Run each workflow with its prompting rules

The official tutorial gives each workflow its own prompting rules — they differ, and using the wrong one is the most common cause of bad output:

T2V — describe the whole scene. Include shot type, scene, action, characters, and camera movement in one flowing paragraph. Describe sounds and dialogue — the model generates synchronized audio with the video. Short prompt? The workflow's prompt enhancer expands it into cinematic instructions automatically.

I2V — describe what happens next, not what is visible. Write the motion, camera movement, and sounds that follow from the input image; do not re-describe what is already in the frame. Anchor the first frame explicitly when continuing a series: phrasing like "Use the provided start image as the first frame."

FLF2V — describe the transition. Write what happens between the two frames, including camera movement and audio. Keep the frames in the same aspect ratio for a smooth interpolation.

What you can do with these workflows

Multishot: the model generates connected shots holding identity, environment, lighting, voice, and style across cuts — a model capability, not a separate workflow. Prompt as a shot list with marked cuts. See our multishot guide.
Native 4K HDR: the official tutorial notes native 4K HDR output at up to 50 FPS, with a RAW workflow for professional finishing pipelines.
Audio: every workflow generates synchronized audio-video; describe dialogue and sound in the prompt.

FAQ

Why do my model downloads fail? The repository is gated. Accept the license on the LTX-2.5 HF page and wait for approval. In ltx-pipelines, also make sure your HF token has the "read gated repos" scope.

Which LTX-2.5 files do I need in ComfyUI? The official T2V/I2V set: int8-convrot distilled transformer, both text encoders, video VAE, audio VAE, and the latent spatial upscaler. FLF2V skips the upscaler.

Why does the workflow use int8 and not the bf16 transformer? The comfy-int8-convrot build is the official ComfyUI path — smaller and designed for consumer VRAM. The bf16 checkpoints are for the ltx-pipelines Python path.

Do I need the Gemma 4 stock encoder too? Yes — the official workflows load two encoders: the LTX-specific gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot and the gemma4_e2b_it_bf16 from Comfy-Org. Both must be in models/text_encoders/.

Does LTX-2.5 work in ComfyUI Cloud? The official tutorial links Cloud runs for all three workflows ("Run on Comfy Cloud"), with the note that Cloud updates after the ComfyUI stable release.

Bottom line

LTX-2.5 in ComfyUI is three native workflows, six files, and one gate: request access to the gated HF repo first, then update ComfyUI, load a template, and place each file in its exact folder — two text encoders included. Then prompt per workflow: whole scene for T2V, what happens next for I2V, the transition for FLF2V. Not ready to manage the local stack yet? Try LTX 2.5 in the browser at ltx23.app — no gated repo, no folder layout.

Sources

Verified August 13, 2026 against primary documentation:

ComfyUI official LTX-2.5 tutorial — three workflows, file list and storage layout, gated-repo warning, prompting tips per workflow
Lightricks/LTX-2.5 — Hugging Face model card — split file pack, ComfyUI-only int8 builds, constraints, license
Lightricks/LTX-2 — GitHub README — gated repo token requirements, int8/FP8 paths, prompting structure
Comfy-Org/gemma-4 — Hugging Face — gemma4_e2b_it_bf16 text encoder required by the official workflows
Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler — Hugging Face — optional detailing IC-LoRA for DFR refinement

Community 8 GB VRAM reports are preliminary third-party claims as of one day after release; verify on your own hardware.