LTX-2.3 vs Wan 2.2: historical benchmark and current options

Magic Hour publishes this historical benchmark and supplied the endpoint and credit meter used for both runs. The retained prompt, outputs, timings, and costs support the narrow comparison below; it should not be generalized beyond the tested LTX-2.3 and Wan 2.2 configurations.

Which of the two should you use?

For a new project, do not choose between these two versions from this benchmark alone. LTX-2.3 and Wan 2.2 remain reproducible historical checkpoints, while LTX 2.5 and MiniMax H3 are the current open-weight starting points in this market. If an existing pipeline is pinned to LTX-2.3 or Wan 2.2, use the measured evidence below and verify the exact checkpoint license, runtime and host before changing it.

Here is what the two runs returned, same prompt, same endpoint:

Cost per clip uses the credit-pack rate of $0.003 per credit. At the usage-based Starter rate the same clip comes to about $0.22.

Both jobs used the same prompt ("A red fox trotting through fresh snow in a birch forest at golden hour, camera tracking low alongside it") at 720p and 5 seconds, on one endpoint and one credit meter. You can run either model locally from the published weights instead. The prompt action below provides a starting brief for a currently available model. It does not reproduce either historical run unless the checkpoint, host, duration and resolution match the recorded setup.

What did the historical runs show?

ltx-2.3 finished between 27 and 43 seconds after job creation. wan-2.2 finished between 232 and 247 seconds. That is a 5.4x gap at the narrowest reading of those windows, roughly 6x at the midpoints.

The outputs differ more than the request did. wan-2.2 returned 1280x720 at 16 fps in a 2.78 MB file; ltx-2.3 returned 1296x720 at 24 fps in an 8.94 MB file. Both jobs cost 240 credits, which works out to 48 credits per second of 720p video on either model. The raw create and status responses, the full run log, and both output clips are in the public benchmark repo.

Local measurements show the same ordering. A three-reviewer head-to-head on an RTX 4090 timed Wan 2.2 14B at 4 minutes 20 seconds per 5-second clip and LTX at 1 minute 30, though its LTX baseline was the older LTX-Video 0.9.5, so the ordering transfers and the exact ratio doesn't. In the same head-to-head's three-reviewer scoring, Wan 2.2 took the highest motion-stability score of the open models, 8.7 against LTX-Video's 7.9.

What did contemporaneous users report?

People who run both models locally report the same split our runs measured. One 5090 owner who runs both: "Does Wan have better motion, anatomy, and quality? Absolutely. But when I can make three and a half videos in LTXV in the time it takes to make one in Wan 2.2 thats a significant advantage." Their timings: 9 minutes for a 10-second 720p Wan clip at FP8, 2.5 minutes for the same clip on LTX at BF16, at 24 fps with sound.

The specific Wan strengths people name:

• Face and body consistency in image-to-video; LTX "really seems to struggle with not totally changing subjects in i2v"
• Anatomy: eyes and eyelids degrade on LTX with multiple people in a scene
• A deeper LoRA and checkpoint ecosystem from Wan's longer run as the community default

The specific LTX strengths:

• Speed, in every account that mentions it
• Synchronized audio in the same pass, which several users call the deciding feature on its own
• Longer single-shot clips: one user renders 20-second full-HD generations with audio, "impossible with wan unless using very complex workflows"
• Low-VRAM viability: a 6 GB user runs 720p 10-second clips on LTX-2.3 where Wan capped them at 480p and 5 seconds

The recurring LTX complaint is prompt adherence: "Prompt adherance is very poor", "just too unpredictable". The fix users converge on is rewriting the prompt with an LLM before generation.

A second thread adds the trajectory point: Wan's open releases stopped at 2.2, so its ecosystem is mature but frozen, while LTX keeps getting open releases and new LoRAs. One user with over 2,000 hours on Wan says LTX "has indeed, at this point, replaced WAN for me."

How were these releases licensed?

For a video model, open weights and open source are different claims. Open weights means the checkpoint is downloadable: you can fetch the model and run it. Open source means the license lets you use the weights, modify them, and ship a commercial product on them without a discretionary approval, a territory restriction, or a revenue trigger.

Apache 2.0 clears that bar. A community license with a $10 million revenue threshold or a jurisdiction exclusion does not. By that test, of the leading open-weights families (Wan, Hunyuan, LTX), only Wan 2.2 is open source.

Wan 2.2's license is the simple one: Apache 2.0, and the README adds that Alibaba claims no rights over generated content.

The LTX-2 Community License requires "Commercial Entities", defined as those with annual revenues of at least $10,000,000, to obtain a paid commercial license from Lightricks, and calls commercial use outside that agreement "strictly prohibited". Lightricks' own FAQ describes LTX-2.3 as "free and open source to use under its license"; the license text above is the binding document.

What hardware did these releases require?

Three open families matter as of July 2026: Wan 2.2, LTX-2/2.3, and HunyuanVideo-1.5. Mochi 1, CogVideoX, and Open-Sora came before. Vendors rarely state hardware requirements upfront, so here they are per model:

Wan 2.2's 14B model is a two-expert mixture-of-experts: 27B parameters total, 14B active per step. It runs in 6 to 8 GB with GGUF quantization and the T5 text encoder offloaded to CPU, needs roughly 54 to 65 GB at FP16, and the repo's own single-GPU command asks for at least 80 GB. The T5-XXL text encoder, 9.4B parameters on its own, dominates the VRAM budget.

The smaller Wan TI2V-5B generates 720p at 24 fps on a 24 GB RTX 4090, a 5-second clip in under 9 minutes. Wan 2.1's 1.3B model, at 8.19 GB, is still the lowest-VRAM option in the family.

LTX-2.3 is a 22B model that generates synchronized video and audio in one pass, shipped as a full dev checkpoint and an 8-step distilled one, with just over 2 million Hugging Face downloads last month. The base LTX-2 is a 19B variant with fp8 and nvfp4 checkpoints. Lightricks says it runs locally on a single RTX 5090.

HunyuanVideo-1.5 is the lightest of the three current families: 8.3B parameters, a 14 GB minimum with model offloading enabled, and a 75-second 480p generation on a single RTX 4090 with the step-distilled checkpoint.

Three cautions before you download anything:

• Pages offering "Wan 2.7 open weights" are SEO fabrications; official Wan open weights stop at 2.2.
• Comfortable VRAM sits higher than the marketing suggests: "all the various video models really want an 80+ gig vram card, to run comfortably". The low-VRAM figures come from community projects like Wan2GP, which gets Wan running on as little as 6 GB.
• AMD owners report ROCm video generation as buggy and OOM-prone, so "runs locally" means NVIDIA in practice.

How were LTX-2.3 and Wan 2.2 run?

The benchmark used a unified hosted API. Magic Hour's video API handled the recorded requests with one request shape and credit balance. The model names and host settings retained below describe that historical catalog; they are not a current availability list. The two requests differed in their model field.

The original article retained this OpenAPI enum alongside the two requests. Keep it as an archival request example, not the full current catalog:

"model": { "default": "default", "type": "string", "enum": [ "default", "ltx-2", "ltx-2.3", "wan-2.2", "seedance-1.5", "seedance-2.0", "seedance-2.0-mini", "kling-2.5", "kling-3.0", "veo3.1", "veo3.1-lite", "sora-2", "kling-1.6", "seedance", "kling-2.5-audio", "veo3.1-audio" ],

The image-to-video product animates a still rather than a written scene. For a new integration, use the current image-to-video endpoint and its model-specific inputs instead of copying the archived enum.

The following table preserves the historical Magic Hour host configuration for resolution, duration and audio. These are host-specific entries from the original article, not universal limits of Google, Kling or the downloadable model weights.

In that archived table, LTX-2.3 was listed up to 30 seconds and Wan 2.2 up to 15, with no audio for Wan 2.2. Do not apply the table’s old model list or maxima to a new host or a current replacement release.

The recorded jobs followed the asynchronous create, status-poll and download workflow. Retain the original request, responses and downloaded MP4 when reproducing the benchmark. A completed job is not by itself proof that the clip passed editorial or product-quality review.

For a new integration, select an explicit current model and record what actually ran. The current endpoint documentation identifies LTX-2.5 as the free-tier default; default can change with the recommendation. A legacy model name is not a checkpoint pin: replacement routing can upgrade an old identifier. To reproduce LTX-2.3, retain its original weights and runtime or obtain an explicit version guarantee from the host.

What does the hosted path not get you?

The hosted catalog has expanded beyond the two families in the old table. The current model reference documents LTX-2.5, MiniMax H3 and Wan 3.0 alongside Wan 2.2. A hosted model’s input controls, duration and licensing path may differ from running public weights; verify the exact endpoint for the job you need.

Choose the deployment path from the controls the project actually requires:

• An exact checkpoint and runtime when historical reproduction matters.
• LoRAs, quantization, seeds and inference steps when the workflow requires those controls; verify that the chosen runtime or endpoint exposes them.
• The selected host’s supported resolution, duration and audio. The current model reference still lists Wan 2.2 without generated audio.

Separate the benchmark's historical dollar assumptions from the rate quoted for a new API job:

The original benchmark's $0.72-per-clip figure used a credit-pack assumption of $3 per 1,000 credits. Its approximately $0.22 usage-based estimate used a different billing assumption. These are the article's historical cost calculations, not a new quote. Check current API billing and the model credit table for a new job. Free trials and signup offers are account-specific; this guide does not promise 400 signup credits or 100 recurring daily credits.

Open Magic Hour to review the current account offer. No coupon discount is assumed in the cost comparison.

Self-host instead if you need LoRA fine-tuning, one of the unserved models, or parameter-level control. The hardware for that path is in the VRAM table above. For the broader hosted-tool market, see our guide to AI video generation tools for creators and studios.

What should you evaluate now?

For a new evaluation, compare LTX 2.5 and MiniMax H3 using the same source assets, prompt intent, duration, resolution and acceptance rubric. LTX 2.5 is the current Lightricks model family. MiniMax H3 provides community-licensed base checkpoints for first/last-frame and multimodal-reference workflows; its official repository also documents hosted components that are not part of the downloadable base release.

Retain Wan 2.2 or LTX-2.3 when compatibility, an existing fine-tune, known hardware behavior or benchmark reproducibility matters. Treat the figures in this article as results from the recorded historical setup, not forecasts for a different checkpoint, host, GPU or billing plan.

If you prefer a hosted route, compare the current video AI API options and record the exact model behind each endpoint. A hosted provider can update a model independently of the public weights.