LTX 2.3 vs Wan 2.2: Specs, File Sizes, Licence and VRAM
LTX 2.3 vs Wan 2.2 from primary sources: parameters, audio, exact Hugging Face file sizes, official VRAM statements, ComfyUI routes and the two licences.
Quick answer
LTX 2.3 vs Wan 2.2 comes down to three differences you can check in the files themselves: LTX-2.3 generates sound with the picture and Wan 2.2 does not, LTX-2.3 is one larger file where Wan 2.2's 14B is two experts, and Wan 2.2 is Apache 2.0 while LTX-2.3 is under Lightricks' own community licence with a revenue threshold.
| Question | LTX-2.3 | Wan 2.2 |
|---|---|---|
| QuestionWho makes it | LTX-2.3Lightricks | Wan 2.2Alibaba's Wan team |
| QuestionOpen weights released | LTX-2.3March 2026 | Wan 2.2July 28, 2025 |
| QuestionAudio | LTX-2.3Generated with the video, in one model | Wan 2.2None in T2V, I2V or TI2V; S2V takes audio as input |
| QuestionMain ComfyUI file | LTX-2.329.15 GB FP8 checkpoint (template) | Wan 2.2Two 14.29 GB FP8 experts (14B), or one 10.00 GB FP16 file (5B) |
| QuestionLicence | LTX-2.3LTX-2 Community License; paid licence at $10M revenue | Wan 2.2Apache 2.0 |
| QuestionOfficial VRAM line | LTX-2.3Lightricks: 32GB+ for its nodes; 12GB in one summary | Wan 2.2ComfyUI docs: the 5B "should fit well on 8GB vram" |
We have not run either model, and we have not run a side-by-side test. Every file size below was read from the Hugging Face API on 2026-10-09. Every speed, quality or VRAM figure is somebody else's and says whose.
Which LTX this page compares
"LTX 2.3" is one point in a family that keeps moving, and search results mix the versions. Lightricks' own sources give this history:
| Version | Parameters | Open weights | Status on 2026-10-09 |
|---|---|---|---|
| VersionLTX-2 | ParametersAbout 19B | Open weightsJanuary 5, 2026 (API launch October 23, 2025) | Status on 2026-10-09Superseded |
| VersionLTX-2.3 | Parameters22B per Lightricks' model page and file names | Open weightsMarch 2026; Hugging Face repo created March 4, UTC | Status on 2026-10-09Labelled "Legacy" in the LTX-2 GitHub README |
| VersionLTX-2.5 | ParametersSee our LTX page | Open weightsAugust 11, 2026 | Status on 2026-10-09Current release |
There is no public LTX-2.4. This page compares LTX-2.3, the 22B release, with Wan 2.2. For LTX 2.3 vs 2.5, the newer model, its int8 files and the one measured 12GB run we found are on our LTX 12GB workflow page. Lightricks' compatibility page says to keep LTX-2.5 and LTX-2.3 workflows and files separate.
Two dates disagree inside Lightricks' own pages. The LTX-2.3 release blog shows February 23, 2026; the newsroom post is dated March 5, 2026, and the Hugging Face repo appeared the night before. The newsroom also calls it a 20.9-billion-parameter model, while the model page and every file name say 22B.
Side by side
| Spec | LTX-2.3 | Wan 2.2 |
|---|---|---|
| SpecParameters | LTX-2.322B (Lightricks model page); 20.9B (Lightricks newsroom) | Wan 2.2A14B: two experts, 27B total, 14B active per step. TI2V-5B: 5B dense |
| SpecArchitecture | LTX-2.3DiT audio-video model; the LTX-2 paper describes a 14B video stream and a 5B audio stream | Wan 2.2Mixture-of-experts diffusion: a high-noise and a low-noise expert |
| SpecText encoder | LTX-2.3Gemma 3 12B | Wan 2.2UMT5-XXL |
| SpecResolution, open weights | LTX-2.3No maximum stated; width and height divisible by 32. Native portrait up to 1080×1920 (blog) | Wan 2.2A14B: 1280×720, 720×1280, 832×480, 480×832. 5B: 1280×704 or 704×1280 |
| SpecFrames and fps, open weights | LTX-2.3Frame count divisible by 8, plus 1. ComfyUI template default: 25 fps, 5 seconds | Wan 2.2Script defaults: 81 frames at 16 fps; the 5B 121 frames at 24 fps |
| SpecHosted API (not local) | LTX-2.3LTX API: up to 4K (3840×2160) at 24 to 50 fps, up to 20 seconds on the fast tier | Wan 2.2Not covered here |
| SpecDistilled or fast variant | LTX-2.3Lightricks distilled checkpoints: 8 steps, CFG 1 | Wan 2.2Third-party lightx2v 4-step LoRAs, one per expert |
| SpecComfyUI route | LTX-2.3Native core templates, or Lightricks' ComfyUI-LTXVideo nodes | Wan 2.2Native core templates |
| SpecLicence | LTX-2.3LTX-2 Community License; entities at $10M+ annual revenue need a paid licence | Wan 2.2Apache 2.0 |
The 4K and 20-second figures belong to Lightricks' hosted API, documented separately from the open weights. The open-weights model card does not state a maximum resolution or duration.
LTX-2.3 files and sizes
Sizes are decimal gigabytes computed from the bytes the Hugging Face API returned. Wan 2.2's full list, with totals, is on our Wan 2.2 VRAM page.
| Repository | File | Bytes | Size |
|---|---|---|---|
RepositoryLightricks/ | Fileltx- | Bytes46,149,344,974 | Size46.15 GB |
RepositoryLightricks/ | Fileltx- | Bytes46,149,345,334 | Size46.15 GB |
RepositoryLightricks/ | Fileltx- | Bytes7,605,507,256 | Size7.61 GB |
RepositoryLightricks/ | Fileltx- | Bytes995,743,560 | Size1.00 GB |
RepositoryLightricks/ | Fileltx- | Bytes29,145,431,166 | Size29.15 GB |
RepositoryLightricks/ | Fileltx- | Bytes29,531,884,062 | Size29.53 GB |
RepositoryLightricks/ | Fileltx- | Bytes21,701,367,310 | Size21.70 GB |
RepositoryComfy- | Filegemma_ | Bytes9,447,702,218 | Size9.45 GB |
RepositoryComfy- | Filegemma- | Bytes628,203,616 | Size0.63 GB |
RepositoryComfy- | Fileltx_ | Bytes2,741,024,390 | Size2.74 GB |
Each LTX-2.3 checkpoint is a single file that bundles the transformer, the video and audio VAEs and the text projection, so 29.15 GB is the whole model except the text encoder. The older ltx- (46,149,345,038 bytes) is still in the repo next to the 1.1 build.
The ComfyUI LTX-2.3 text-to-video template loads the FP8 dev checkpoint, the FP4 Gemma encoder, the Gemma LoRA, the distilled 1.1 LoRA at strength 0.5 and the x2 upscaler. Those five files total 42,958,104,950 bytes, 42.96 GB. Wan 2.2's 14B text-to-video template, with its 4-step LoRAs, is 38.03 GB; the 5B template is 18.14 GB, both per our Wan page.
LTX 2.3 GGUF
GGUF builds come from third parties, not Lightricks:
| Repository | File | Bytes | Size |
|---|---|---|---|
Repositoryunsloth/ | Filedistilled 1.1, Q3_K_M | Bytes10,629,508,128 | Size10.63 GB |
Repositoryunsloth/ | Filedistilled 1.1, Q4_K_M | Bytes14,194,441,248 | Size14.19 GB |
Repositoryunsloth/ | Filedistilled 1.1, Q8_0 | Bytes22,755,540,000 | Size22.76 GB |
RepositoryQuantStack/ | Filedistilled 1.1, Q4_K_M | Bytes17,763,015,328 | Size17.76 GB |
Unlike the single-file checkpoint, the unsloth GGUF splits the parts out: the embeddings connector is another 2.31 GB, the video VAE 1.45 GB and the audio VAE 0.36 GB. For comparison, one Wan 2.2 14B expert at Q4_K_M is 9.65 GB, and the 14B needs two.
LTX 2.3 ComfyUI routes
- Native templates. ComfyUI's LTX-2.3 tutorial lists six workflows, including
video_andltx2_ 3_ t2v .json video_, and says no custom nodes are required. Its defaults are 1280×720, 25 fps, 5 seconds, two stages, CFG 1. The page gives no VRAM figure.ltx2_ 3_ i2v .json - Lightricks ComfyUI-LTXVideo. Lightricks' own nodes, with LTX-2.3 workflows under
example_. Its README asks for a CUDA GPU with 32GB+ VRAM and 100GB+ of free disk.workflows/ 2.3/ - GGUF. The files above load through ComfyUI-GGUF, as on our LTX 12GB page.
Wan 2.2 is in ComfyUI's core templates too, since July 28, 2025. Where each Wan file goes is on our Wan 2.2 models folder page.
LTX 2.3 prompting guide
The LTX-2.3 model card points to Lightricks' guide at ltx. The newer docs.ltx.io prompting guide is written for LTX-2.5. Both are Lightricks' documents; we have not tested their advice.
Speed and quality: who claimed what
We have run no benchmark of either model. These are the published claims, attributed:
| Claim | Who | Which model | Measured? |
|---|---|---|---|
| Claim1.22 s vs 22.30 s per step, "approximately 18× faster" | WhoLTX-2 paper, Lightricks | Which modelLTX-2 19B vs Wan 2.2-14B, H100 | Measured?By Lightricks; LTX-2, not LTX-2.3 |
| ClaimRanked above Wan 2.2-14B on Artificial Analysis | WhoLTX-2 paper, Lightricks | Which modelLTX-2, as of November 6, 2025 | Measured?Third-party leaderboard, cited by Lightricks |
| ClaimOne-fifth to one-tenth the compute of leading models | WhoLightricks newsroom | Which modelLTX-2.3 | Measured?No method given |
| ClaimWan 2.2 beats leading closed models on Wan-Bench 2.0 | WhoWan README | Which modelWan 2.2; LTX not named | Measured?Wan's own benchmark |
| ClaimLTX-2.3 10–14× faster, once about 18× | WhoWaveSpeed blog author | Which modelBoth, RTX 4090, ComfyUI | Measured?Author's own runs, settings not all stated |
| Claim27–43 s vs 232–247 s for one 720p, 5-second clip | WhoMagic Hour | Which modelBoth, on its own hosted API | Measured?Measured on one prompt, cloud hardware |
None of these measures LTX-2.3 and Wan 2.2 on a consumer card with the same settings and published VRAM. Quality verdicts on those pages are the authors' opinions; we repeat none of them.
Why figures online disagree
- Two licences called one. Some comparison pages list LTX-2.3 as Apache 2.0. Lightricks says its model licence is not Apache 2.0. The Apache licence Lightricks announced in March 2026 is for LTX Desktop, the app.
- Wan 2.1 in Wan 2.2 tables. A "1.3B lite" Wan is a Wan 2.1 model. Wan 2.2 ships 5B and A14B.
- API specs as model specs. The 4K, 50 fps and 20-second figures are the hosted LTX API's.
- Lightricks' own VRAM lines conflict. Its docs and ComfyUI nodes say 32GB+ minimum. Its ltx.io model summary says the distilled and FP8 variants run on as little as 12GB. Neither gives a measured peak.
- The 18× is LTX-2, not 2.3. It comes from the LTX-2 paper, on an H100.
Which to pick by card and RAM
These are starting points read off the file sizes and the statements above, not test results. Check your own numbers on the system requirements checker; neither LTX-2.3 nor Wan 2.2 is one of its presets yet, only MiniMax H3.
| Your machine | Starting point |
|---|---|
| Your machine8GB card | Starting pointWan 2.2 5B, the only one of these with an official 8GB statement. No LTX-2.3 file fits; the smallest GGUF above is 10.63 GB. |
| Your machine12GB card, 32GB RAM | Starting pointWan 2.2 5B, or the 14B as Q4_K_M experts at 9.65 GB each. LTX-2.3 needs part of every file in system RAM. |
| Your machine12GB card, 48GB+ RAM | Starting pointEither. The one measured 12GB LTX run we found was LTX-2.5 with 48GB RAM, peaking at 45 GB RAM. Nobody has published a 2.3 run. |
| Your machine16GB to 24GB card | Starting pointEither. LTX-2.3 unsloth Q4_K_M fits on 16GB by size; the FP8 checkpoint at 29.15 GB still streams. Wan 14B FP8 experts are 14.29 GB. |
| Your machine32GB+ card | Starting pointEither, at Lightricks' stated minimum for LTX. |
| Your machineYou need sound in the clip | Starting pointLTX-2.3. Wan 2.2's text and image models make silent video. |
| Your machineRevenue at or above $10M | Starting pointRead the LTX licence first, or use Wan 2.2 under Apache 2.0. |
For a 12GB card specifically, our RTX 3060 12GB video model page compares the options. Our only own measurement is MiniMax H3, not LTX or Wan: on an RTX 3060 it peaked at 11,649 MiB of VRAM and 43,587 MiB of system RAM.
What nobody has published
- A measured LTX-2.3 vs Wan 2.2 run on one consumer GPU, same resolution and length, with peak VRAM and RAM.
- A Lightricks VRAM figure for the native ComfyUI LTX-2.3 templates.
- An official maximum resolution or duration for the LTX-2.3 open weights.
- Any LTX-2.3 vs LTX-2.5 quality comparison from Lightricks beyond its release notes.
Licence and downloads
LTX-2.3: Lightricks' LICENSE file says the LTX-2 Community License of January 5, 2026 covers LTX-2 releases, including LTX-2.3 until August 11, 2026, and the LTX-2.x Community License covers LTX-2.5 onward. The LTX-2.3 model card's metadata points to the first, while its body links the second. Both say entities with annual revenue of $10,000,000 or more need a paid commercial licence, and both restrict use in products that compete with Lightricks. Read the licence linked from the repository before building on it.
Wan 2.2: the Wan README says it is licensed under Apache 2.0, claims no rights over generated content, and restricts harmful uses.
We do not host any of these files. Download them from the repositories named above.
GenVidKit is an independent guide. It is not affiliated with Lightricks, Alibaba, the Wan team, Comfy Org, unsloth, QuantStack, lightx2v or Hugging Face.
Sources
All read on 2026-10-09.
- Lightricks/LTX-2.3, LTX-2.3-fp8 and LTX-2.3-nvfp4 on Hugging Face — byte counts, repo dates, model card tips, licence metadata, prompting-guide link.
- Comfy-Org/ltx-2 and Comfy-Org/ltx-2.3 — Gemma encoder and LoRA byte counts.
- unsloth/LTX-2.3-GGUF and QuantStack/LTX-2.3-GGUF — GGUF byte counts.
- Lightricks/LTX-2 on GitHub — the "Legacy" label, the
LICENSE,LICENSE-and2 LICENSE-files,2_ x MODELS-.LTX- 2.3 .md - Lightricks/ComfyUI-LTXVideo — the 32GB+ prerequisite and the 2.3 example workflows.
- LTX llm-info — version history, 19B and 22B, the 12GB line, "not Apache 2.0".
- LTX-2.3 release blog — the February 23 date and the 1080×1920 portrait claim.
- LTX-2.3 newsroom post and LTX-2 open-source newsroom post — March 5 and January 5 dates, 20.9B, the compute claim, the LTX Desktop Apache note; the newsroom index dates the LTX-2 API launch to October 23, 2025.
- LTX docs: system requirements and LTX-2.3 API model page — the 32GB minimum, the API resolutions, fps and durations.
- LTX-2 paper, arXiv 2601.03233 — the 14B plus 5B streams, the per-step timing against Wan 2.2, the Artificial Analysis ranking.
- LTX-2.3 ComfyUI workflow examples and Comfy-Org/workflow_templates — the six native workflows and which files the templates load.
- Wan-Video/Wan2.2 on GitHub — release dates, the 27B and 14B-active design, sizes, script defaults, Wan-Bench claim, licence.
- Wan2.2 native workflow, ComfyUI documentation — the 8GB statement for the 5B.
- WaveSpeed, Magic Hour and CrePal — the third-party claims above, and the Apache 2.0 and 1.3B errors.