Every "LTX 2.3 vs Wan 2.2" post you will find is a quality shootout: same prompt, two clips, pick a winner. That comparison is nearly useless, because it answers a question that changes every month and ignores three that do not.
The three durable questions are: does it generate audio, what does the license let you do, and will it run on the GPU you have? Those determine whether a model fits your project. Frame-by-frame preference does not survive the next point release.
This comparison is built only on primary sources: the LTX-2.3 model card and LTX-2 repository on one side, the Wan2.2-T2V-A14B model card and Wan2.2 repository on the other. Checked July 2026.
The spec sheet, from official cards only
| LTX 2.3 | Wan 2.2 (T2V-A14B) | |
|---|---|---|
| Architecture | 22B dense, DiT-based audio-video foundation model | Mixture-of-Experts, two experts, ~27B total, 14B active per step |
| Audio | Synchronized audio generated with video | Not described as an audio-generating model |
| Checkpoint size | 46.1 GB (bf16) | Varies by variant |
| Documented VRAM | No published inference minimum; FP8 + CPU/disk offload documented; training 80 GB+ (32 GB with INT8) | Single-GPU inference stated as requiring at least 80 GB VRAM, reducible via offload and dtype flags |
| Resolution notes | Width/height divisible by 32; frame count 8k+1 | 5s videos at 480P and 720P; the separate TI2V-5B variant does 720P at 24 fps |
| License | LTX-2 Community License Agreement | Apache 2.0 |
| Released | LTX-2 open-sourced Jan 2026; LTX 2.3 in Mar 2026 | Inference code and weights released Jul 28, 2025 |
| ComfyUI | Six native templates, no custom nodes | Widely supported |
Two entries in that table decide most real projects. Let's take them in order.
1. Audio is the actual dividing line
LTX-2 is described by Lightricks as the first DiT-based audio-video foundation model — one model producing synchronized audio and video together. Wan's T2V-A14B card describes video generation and makes no audio-generation claim.
This is not a feature-checklist difference. It changes your entire pipeline.
If you need talking characters, ambient sound, or music-synced motion, an audio-capable model means one generation step. The alternative is a chain: generate video, generate or record audio separately, then drive lip-sync with a third tool, then fix drift. Each hop is a place for sync to break.
LTX 2.3's audio support is deep enough that it has its own dedicated tooling — a LipDubPipeline for matching lip movement to new audio against a reference clip, an A2VidPipelineTwoStage for generating video conditioned on an input audio track, and even a T2AOneStagePipeline that generates audio only, with no video branch at all. Lightricks also publishes audio-focused training configs (t2a_lora, v2a_lora, audio extension and inpainting).
If your output has sound in it, this is the whole comparison. If your output is silent B-roll, it barely matters and you should weigh the other two factors.
2. Licensing: the difference that shows up in legal review
This is where the two projects genuinely diverge, and it is under-discussed because it is boring right up until it is not.
Wan 2.2 is Apache 2.0 — a standard permissive license, no revenue threshold, with restrictions on harmful content generation.
LTX 2.3 ships under the LTX-2 Community License Agreement, and it contains a revenue trigger. The license text states that entities with annual revenues of at least $10,000,000 are defined as "Commercial Entities" and are required to obtain a paid commercial use license to use LTX-2 or its derivatives. The agreement also requires that derivatives be distributed under the same agreement and that attribution notices be preserved.
Practically:
- Individual creators, freelancers, and small studios — below the threshold, the community license is a non-issue.
- Startups approaching $10M in revenue — put this on the roadmap review, not the day it becomes true.
- Enterprises — this is a procurement conversation with Lightricks before deployment, not after.
- Anyone redistributing a fine-tune — derivatives inherit the agreement. Read the copy in the repo yourself; a blog summary is not legal advice, including this one.
Apache 2.0 has none of that machinery. If your organization's legal posture is "permissive licenses only, no negotiated agreements," that fact alone may decide this comparison regardless of output quality.
3. Single-GPU feasibility: neither is a small model
Both are heavy, but they document it differently, and that difference is itself informative.
Wan 2.2 T2V-A14B states single-GPU inference requires at least 80 GB VRAM, reducible with offload and dtype conversion flags. Its MoE design activates ~14B parameters per step out of ~27B total — a real efficiency win at inference time, but the weights still have to live somewhere.
LTX 2.3 publishes no inference minimum. What it documents instead is a full set of reduction levers: --quantization fp8-cast (any FP8-capable GPU) or fp8-scaled-mm (Hopper+), --offload {cpu,disk}, --max-batch-size, gradient estimation to cut 40 steps to 20–30, and ComfyUI FP8 checkpoint builds.
Neither project is promising you a 12 GB experience. The honest framing: both are datacenter-shaped models with consumer escape hatches. LTX 2.3's escape hatches are more thoroughly documented, and the ComfyUI FP8 route is the most accessible path either project offers.
Also worth noting for anyone comparing on speed: LTX 2.3's DistilledPipeline runs 8 steps in stage 1 and 4 in stage 2, with no guidance required. That is a fundamentally different iteration loop from a standard guided sampler, and it is often what people are actually reacting to when they call one model "faster."
Don't want to benchmark two 46 GB models to decide? You can generate with LTX 2.3 right now at ltx23.app — including audio — and see whether it clears your bar before you download anything.
Where each one is the obvious pick
Choose LTX 2.3 when:
- Your output has dialogue, ambient sound, or music sync
- You want lip-dubbing, audio-driven video, or voice-cloned identity in one stack
- You want native ComfyUI templates with no custom nodes
- You want region-level repair (
RetakePipeline) instead of rerolling clips - You need HDR linear-float output for grading
- Your revenue is comfortably below the license threshold
Choose Wan 2.2 when:
- Apache 2.0 is a hard requirement from legal or your open-source policy
- Your output is silent and audio buys you nothing
- Your team already has Wan tooling and LoRAs in production
- MoE inference efficiency suits your serving setup
Honestly? Many people should run both. They are free to download, and the decision cost is disk space rather than money. Just make the licensing decision deliberately — that one is not reversible after you have shipped.
What this comparison deliberately does not claim
No quality scores. Wikipedia's LTX-2 entry cites Artificial Analysis rankings at LTX-2's release — third for image-to-video, seventh for text-to-video — but that is a secondary source, a snapshot in time, and about LTX-2 rather than 2.3. Treating it as a current verdict would be exactly the mistake this article is arguing against.
Head-to-head visual quality also depends on your prompt style, your resolution, your step count, and which checkpoint variant you loaded. Anyone publishing a clean "X beats Y" on video quality is showing you their settings as much as the models. Run your own prompts on your own hardware — that is the only benchmark whose result you can act on.
FAQ
Does Wan 2.2 generate audio? Its official T2V-A14B card describes video generation and does not claim audio generation. LTX 2.3 generates synchronized audio and video in one model.
Is LTX 2.3 open source? Open weights under the LTX-2 Community License Agreement, which includes a paid-license requirement for entities at or above $10M annual revenue. Wan 2.2 is Apache 2.0.
Which needs less VRAM? Wan states 80 GB+ for single-GPU inference. LTX 2.3 publishes no minimum but documents FP8 and offload paths plus FP8 ComfyUI builds. Neither is a low-VRAM model in its native form.
Which is faster? Different mechanisms: Wan's MoE activates ~14B of ~27B per step; LTX 2.3's distilled path runs 8+4 steps with no guidance. Compare on your hardware.
Can I use both in ComfyUI? Yes. LTX 2.3 has six native templates requiring no custom nodes; Wan 2.2 is widely supported too.
Bottom line
Audio, license, hardware — decide in that order and the choice usually makes itself. LTX 2.3 is the pick when sound is part of the output and you are below the commercial-license threshold. Wan 2.2 is the pick when Apache 2.0 is non-negotiable or your work is silent.
Test LTX 2.3's audio-plus-video output in your browser at ltx23.app before you spend a weekend downloading both.
Sources
Verified July 2026 against primary documentation:
- Lightricks/LTX-2.3 — Hugging Face model card — variants, sizes, constraints, license name
- LTX-2 Community License Agreement (repository copy) — $10,000,000 revenue threshold, derivative and attribution terms
- Lightricks/LTX-2 — official GitHub repository — audio-video foundation model description, pipelines, optimization flags
- LTX-2 available pipelines documentation — LipDub, A2Vid, T2A, Retake, HDR pipelines and step counts
- LTX-2 optimization documentation — FP8 policies, gradient estimation
- Wan-AI/Wan2.2-T2V-A14B — Hugging Face model card — MoE design, active parameters, 480P/720P, 80 GB VRAM, Apache 2.0, release date
- Wan-Video/Wan2.2 — official GitHub repository — inference code and variants
- ComfyUI official LTX-2.3 tutorial — native templates and FP8 builds
- ltx-trainer documentation — training hardware requirements
- LTX-2 technical paper (arXiv 2601.03233) — architecture
- Wikipedia: LTX-2 — release timeline and third-party benchmark placement (secondary source, cited as such)
Licensing summaries here are not legal advice. Read the license text in each repository before making a commercial deployment decision.
