WAN 2.2 vs LTX 2.3
WAN 2.2 and LTX 2.3 are two of the strongest open-weight video generation models you can download today — and two of the most compared. The problem is that most comparisons mix marketing claims with spec sheets. This guide compares the two from primary sources only: Alibaba's official WAN 2.2 repository, Lightricks' official LTX-2.3 pages, and the model cards on Hugging Face.
TL;DR
- WAN 2.2 (Alibaba Wan-Video team) shipped its open weights and inference code on July 28, 2025. It uses a Mixture-of-Experts (MoE) architecture: the A14B models total 27B parameters but only 14B are active per denoising step, alongside a compact TI2V-5B dense model that generates 720P@24fps and can run on an RTX 4090-class GPU (~24GB VRAM).
- LTX 2.3 (Lightricks) is a 22B DiT-based audio-video foundation model that generates synchronized audio and video in a single model — text-to-video, image-to-video, audio-to-video, and video-to-video — with a rebuilt VAE, a 4× larger text connector, and native 1080×1920 portrait output.
- On H100 GPUs, Lightricks reports the LTX-2 series achieves roughly 18× the throughput of WAN 2.2 14B (vendor-reported, not independently verified).
- Pick LTX 2.3 for audio and video from one model, native vertical video, or fast batch iteration. Pick WAN 2.2 for Apache-2.0 licensing or a lightweight 5B model that fits in 24GB of VRAM.
What is WAN 2.2?
WAN 2.2 is Alibaba's open-source video model family, documented in the official Wan-Video/Wan2.2 repository. The headline innovations, per the repo, are an MoE architecture (high-noise and low-noise experts, ~14B active parameters per step), a high-compression VAE with a total compression ratio of 64 for the 5B model, and training on significantly more data than WAN 2.1 (+65.6% images, +83.2% videos).
The family includes: T2V-A14B and I2V-A14B (text/image-to-video at 480P and 720P), TI2V-5B (a unified text+image-to-video model and one of the fastest 720P@24fps open models), S2V-14B (speech-to-video, announced Aug 26, 2025), and Animate-14B (character animation, announced Sep 19, 2025). Everything is licensed Apache 2.0 and integrated into ComfyUI and Diffusers. On hardware, the A14B text-to-video reference setup wants around 80GB of VRAM, while TI2V-5B is designed for 24GB-class consumer GPUs.
What is LTX 2.3?
LTX 2.3 is Lightricks' open-weight audio-video foundation model, documented on the official LTX-2.3 page and the Lightricks/LTX-2.3 model card. It is a significant update to LTX-2 with improved audio and visual quality and better prompt adherence. The official highlights are a rebuilt VAE (sharper fine detail, cleaner edges), a 4× larger text connector (tighter prompt adherence for complex, multi-subject prompts), cleaner audio (filtered training data plus a new vocoder), and native portrait video up to 1080×1920.
The 22B model ships as ltx-2.3-22b-dev (fully trainable), ltx-2.3-22b-distilled (8 steps, CFG=1) and distilled v1.1, plus LoRAs and spatial/temporal upscalers. It uses a Gemma-3 12B text encoder, and Lightricks provides official pipelines in the LTX-2 GitHub repository, ComfyUI nodes, and a managed API. Licensing is the LTX-2 Community License — free for qualifying organizations under $10M in annual revenue. LTX 2.5 is the newer model in the family; LTX 2.3 remains supported.
WAN 2.2 vs LTX 2.3: side-by-side
| Dimension | WAN 2.2 | LTX 2.3 |
|---|---|---|
| Developer | Alibaba (Wan-Video) | Lightricks |
| Architecture | MoE (A14B: 27B total, 14B active) + dense TI2V-5B | DiT-based audio-video model, 22B |
| Modes | T2V, I2V, TI2V, S2V (speech), Animate | T2V, I2V, A2V (audio), V2V, depth control |
| Audio | Separate S2V-14B model (speech-to-video); base T2V/I2V are video-only | Synchronized audio + video in one model |
| Native portrait | Vertical 704×1280 supported for TI2V-5B | Native 1080×1920 portrait, trained on portrait data |
| Local hardware | TI2V-5B ~24GB VRAM; A14B T2V ~80GB VRAM (ref setup) | Recommended 32GB+ VRAM; fp8/quantized variants available |
| Speed (vendor claim) | — | ~18× throughput vs WAN 2.2 14B on H100 (Lightricks) |
| License | Apache 2.0 | LTX-2 Community License (free < $10M revenue) |
| Ecosystem | ComfyUI, Diffusers, Wan2GP | ComfyUI, LTX-2 pipelines, Diffusers, Wan2GP |
How to decide: LTX 2.3 vs WAN 2.2
There is no universal "better" model — the right choice depends on your output format, hardware, and workflow:
- Audio is the biggest structural difference. WAN 2.2's base video models do not generate sound; synchronized speech-to-video needs the separate S2V-14B model. LTX 2.3 is a joint audio-video model: one checkpoint, one pipeline, and your prompt can produce a clip with a soundtrack, lip-synced speech, or beat-aligned motion.
- Vertical video for social. LTX 2.3 is the first model in its line to support native 1080×1920 portrait output trained on real portrait data — a real advantage for Reels, Shorts, and TikTok. WAN 2.2's TI2V-5B supports a vertical 704×1280 resolution, so it is not a deal-breaker, just lower-res.
- Hardware. If you have a 24GB consumer GPU, WAN 2.2's TI2V-5B is a proven fit. If you have a larger GPU (32GB+ is the recommended local setup for LTX 2.3) and want full 22B quality plus audio, LTX 2.3 is the better match — and quantized fp8/GGUF variants lower the bar.
- Licensing. WAN 2.2 is Apache 2.0. LTX 2.3 uses the LTX-2 Community License, free for organizations under $10M in annual revenue; larger deployments need a commercial license.
- Speed. For batch throughput, Lightricks reports the LTX-2 series at roughly 18× the throughput of WAN 2.2 14B on H100 GPUs — a vendor-reported number, so treat it as directional, not an independent benchmark.
For context: the wider open-source field (Wan 2.1, Hunyuan Video, LTX-Video) is all supported by local launchers like Wan2GP, so you are not locked into one model.
How to run LTX 2.3 locally on Pinokio with Wan2GP
If you want to run LTX 2.3 locally, Wan2GP is a free open-source launcher built "for the GPU poor" that supports Wan 2.1/2.2, LTX-2, Hunyuan Video, LTXV, and more, with low-VRAM profiles and quantized checkpoint formats. It is published on Pinokio (listed under the pinokiofactory wan app). To load a custom LTX 2.3 checkpoint such as the community 10Eros_v1.3_bf16.safetensors file, follow these steps:
- Install and launch Wan2GP via Pinokio. Search "Wan2GP" in the Pinokio app browser, install it, and launch the web UI.
- Download the checkpoint. Grab
10Eros_v1.3_bf16.safetensorsfrom the TenStrip/LTX2.3-10Eros repository on Hugging Face. This repo carries a "Not-For-All-Audiences" notice, so accept the terms before downloading; grab the matching text encoder from the same repo if you are not using Wan2GP's bundled one. - Register it as a custom finetune. In Wan2GP, use the built-in Models/Checkpoints Manager (or the Finetune Manager plugin) to add the
.safetensorsfile as a custom checkpoint under the LTX-2 model family. The official FINETUNES guide shows where files go and how the config maps model components. - Select the model and generate. Switch to the LTX-2 model tab, pick your 10Eros checkpoint, then set the prompt, resolution, and frame count. Remember LTX 2.3's official constraint: width and height must be divisible by 32, and the frame count must be divisible by 8, plus one.
Prefer not to manage GPUs and checkpoints? The hosted LTX 2.3 model handles rendering in the cloud — no GPU or install required — and the free-to-start account covers text-to-video, image-to-video, and audio-to-video. Try LTX 2.3 online at ltx23.app.
FAQ
Is LTX 2.3 better than WAN 2.2? It depends on the task. LTX 2.3 wins for synchronized audio-video in one model, native 1080×1920 portrait, and (per Lightricks) throughput; WAN 2.2 wins on Apache-2.0 licensing and a 5B model that runs in 24GB of VRAM.
Does WAN 2.2 generate audio? Not in its base T2V/I2V models — it ships a separate S2V-14B speech-to-video model. LTX 2.3 generates audio and video together in a single model.
Can I run WAN 2.2 and LTX 2.3 on the same machine? Yes — launchers like Wan2GP and ComfyUI support both model families, so you can install both and switch between them.
How much VRAM does LTX 2.3 need? The recommended local setup is an NVIDIA GPU with 32GB+ VRAM; quantized fp8 and GGUF variants reduce the requirement, and cloud services handle rendering entirely.
Is LTX 2.3 free to use? The weights are open and free for personal and commercial use under the LTX-2 Community License for qualifying organizations under $10M in annual revenue. Larger companies need a commercial license.
Sources
- LTX-2.3 official model page — Lightricks
- Lightricks/LTX-2.3 model card — Hugging Face
- Wan-Video/Wan2.2 official repository — GitHub
- deepbeepmeep/Wan2GP — GitHub
- TenStrip/LTX2.3-10Eros checkpoint — Hugging Face
Scope note: Sources are official vendor pages, model cards, and repository documentation only. The 18× throughput figure is a vendor-reported claim; hardware figures reflect the reference setups documented in the linked repositories. Published 2026-08-25.
