AI Infrastructure

LTX 2.5 in ComfyUI: Files, VRAM, GGUF Builds, Licence and What Changed From 2.3

LTX 2.5 is Lightricks' open-weight 22B model that generates video and sound together. This guide lists every file and its size, the VRAM each variant needs against Lightricks' 32 GB minimum, the ComfyUI templates and folders, the community GGUF builds with their measured quality loss, the official prompt rules, the licence threshold and the differences from LTX 2.3, all read on 14 November 2026.

LTX 2.5 runs in ComfyUI 0.32.0 and later through built-in templates. Lightricks states a minimum of 32 GB of VRAM on an NVIDIA GPU, but its lower-VRAM files are a 21.50 GB INT8 transformer and a 15.37 GB text encoder, and community GGUF builds start at 8.46 GB. The 22B model generates video and synchronised audio, and is free for commercial use below 10 million US dollars in annual revenue.

LTX 2.5 arrived on 11 August 2026. That is the date on its licence, the date Lightricks’ LTX-2 code repository tagged version 1.2.0 with the note “Support for LTX 2.5”, and the date ComfyUI 0.32.0 shipped with the line “Add support for LTX 2.5”. Three months on, Hugging Face counts about 2.8 million downloads of the repository in the past month, and dozens of quantised community builds sit beside it.

We read Lightricks’ model card, repository, licence and documentation, ComfyUI’s tutorials and release notes, and the main community repositories on 14 November 2026, and this page reports what they say. File sizes come from the Hugging Face file listings. Where we estimate, we say so. To set LTX 2.5 beside Wan 2.2, HunyuanVideo 1.5 and the rest, use our comparison of open video models by GPU and licence; this page stays on LTX 2.5 itself.

What LTX 2.5 is

Lightricks describes LTX 2.5 as an open-weight “world model” built for local execution and fine-tuning. In practice it is a 22-billion-parameter diffusion transformer that generates video and audio in the same pass from text, an image, or reference video and audio. The model card lists what is new in this version:

  • Native multishot generation. One run can produce several connected shots that hold character, environment, lighting, voice and style across cuts. Earlier versions produced one continuous shot.
  • Diffusion Fidelity Rendering (DFR). A keyframes-first pipeline that spends compute according to scene complexity.
  • A diffusion video decoder. It replaces the VAE reconstruction stage and targets faces, textures and on-screen text. A lighter convolutional VAE is also published.
  • A custom Gemma 4 12B text encoder. Stock Gemma 4 is not a substitute; the loader checks for Lightricks’ tuned version.
  • A prompt enhancer. A small Gemma 4 E2B model that expands short prompts before encoding.
  • A duration predictor. An optional head that picks the clip length from the prompt.
  • A stronger distilled model. A fixed 8-step schedule at CFG 1.

Two transformers are published. The distilled checkpoint is the one every ComfyUI template uses. The dev checkpoint is the full, trainable model for guided two-stage pipelines and for LoRA training with the LTX-2 Trainer.

Search results also mix in two hosted products, ltx-2-5-fast and ltx-2-5-pro. Those are API models on Lightricks’ servers, not downloadable files.

LTX 2.5 size: every file on Hugging Face

LTX 2.5 ships as a split pack, one .safetensors file per component, in the gated Lightricks/LTX-2.5 repository. You must be signed in and click through the licence before anything downloads.

ComponentFileSize
Distilled transformer, BF16ltx-2.5-22b-distilled-transformer-bf1642.02 GB
Distilled transformer, INT8 ConvRot (ComfyUI only)ltx-2.5-22b-distilled-transformer-comfy-int8-convrot21.50 GB
Distilled transformer, NVFP4ltx-2.5-22b-distilled-transformer-nvfp418.72 GB
Dev transformer, BF16 / INT8 ConvRotltx-2.5-22b-dev-transformer-…42.02 / 21.50 GB
Gemma 4 12B text encoder, BF16 / INT8 ConvRotgemma4-12b-with-proj-ltx-2.5-…26.26 / 15.37 GB
Video VAE, diffusion decoder / convltx-2.5-video-vae-bf16 / -conv-bf161.47 / 1.45 GB
Audio VAE and vocoderltx-2.5-audio-vae-bf160.36 GB
Spatial upscaler ×2 / temporal upscaler ×2ltx-2.5-latent-…-upscaler-x2-bf16-1.01.00 / 0.26 GB
Distilled LoRA for dev workflowsltx-2.5-22b-distilled-lora-450-bf168.90 GB
Duration headltx-2.5-duration-head-bf163.8 MB

Three things follow from the table.

First, the full-precision set is large. Lightricks’ README puts the BF16 transformer, text encoder, two VAEs and the spatial upscaler at roughly 66 GiB.

Second, the lower-VRAM set that Lightricks documents for ComfyUI swaps three files: the INT8 ConvRot transformer, the INT8 ConvRot text encoder and the conv video VAE. With the audio VAE and the upscaler that comes to about 40 GB on disk. The INT8 ConvRot files load only in ComfyUI, not in Lightricks’ Python pipelines.

Third, the NVFP4 transformer is aimed at NVIDIA Blackwell GPUs. The model card says it runs in ComfyUI, or in ltx-pipelines with --quantization nvfp4-prequant and the ltx-kernels package.

The optional prompt enhancer, gemma4_e2b_it, comes from the separate Comfy-Org/gemma-4 repository and is about 5 GB according to ComfyUI’s tutorial. The DFR pipeline also needs a 0.33 GB detailing IC-LoRA from a second gated repository, Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler.

LTX 2.5 VRAM requirements by variant

Lightricks’ system requirements page is short:

MinimumRecommended
GPUNVIDIA with 32 GB or more of VRAMNVIDIA A100 80 GB or H100
System RAM32 GB64 GB or more
Storage100 GB free200 GB or more, on SSD
SoftwareCUDA 12.7 or higher, Python 3.12 or higherCUDA 12.7 or higher

Lightricks is not consistent on this point. The comparison table on its LTX-2.5 product page lists “Min VRAM: 16GB”, with no file or workflow named. The documentation figure is the one to plan against, and the smaller figure is plausible only with the quantised files below.

The arithmetic is simple, and it is the same parameters × bits ÷ 8 sum that our VRAM calculator applies to language models. Twenty-two billion parameters need about 42 GB at 16 bits, 21.5 GB at 8 bits and 14 GB at a little over 5 bits. Activations come on top and grow with resolution and frame count. The text encoder runs once per prompt, and ComfyUI can move it out of VRAM before sampling, so the transformer usually decides which card you need.

VRAMTransformer to tryBasis
8 to 10 GBGGUF Q2_K (8.46 GB) or UD-Q2_K_XL (9.50 GB), encoder offloadedOur estimate from file size; Unsloth measures a large quality drop at 2-bit
12 GBGGUF Q3_K_M (11.30 GB) with offloadingOur estimate from file size
16 GBGGUF UD-Q4_K_XL (14.0 GB)Our estimate; Unsloth’s recommended file for most GPUs
24 GBOfficial INT8 ConvRot (21.50 GB), NVFP4 (18.72 GB) or GGUF Q8_0 (22.78 GB)Our estimate; below Lightricks’ stated minimum
32 GBOfficial lower-VRAM set with room for longer clipsLightricks’ documented minimum
48 GBBF16 transformer (42.02 GB), encoder swapped outOur estimate from file size
80 GBFull BF16 stackLightricks’ recommended A100 80 GB or H100

Every row below 32 GB sits under the vendor’s own specification, so treat it as something to test, not a promise. Offloading also moves the pressure to system memory: the INT8 encoder alone is 15.37 GB, which is why 32 GB of RAM is the stated floor and 64 GB the recommendation.

For the Python pipelines, the model card’s low-VRAM advice is to downcast the BF16 transformer at load time and offload to the CPU:

uv run python -m ltx_pipelines.distilled \
  ...existing flags... \
  --quantization fp8-cast \
  --offload cpu

On Macs, the requirements page lists only NVIDIA. The repository changelog shows that macOS installs resolve to PyTorch’s MPS builds since version 1.3.0, and a community MLX port exists on Hugging Face, which its authors label a research port. Neither is a supported configuration in Lightricks’ documentation. If you are choosing a card for this and for language models, our guide to picking a GPU for local AI covers the 16, 24 and 32 GB options.

LTX 2.5 ComfyUI workflow: set-up steps

Support is native. ComfyUI’s core has loaded LTX 2.5 since 0.32.0 on 11 August 2026; version 0.45.0 was current on 14 November, and 0.44.0 fixed a crash in the LTX 2.5 diffusion VAE decoder, so update before you start. Four official templates cover the common jobs:

TemplateIDStages
Text to Videovideo_ltx2_5_t2vTwo: base, then 2× upscale and refine
Image to Videovideo_ltx2_5_i2vTwo
First-Frame / Last-Framevideo_ltx2_5_flf2vOne; no upscaler
DFRCovered in ComfyUI’s DFR tutorialThree, the last optional
  1. Update ComfyUI. ComfyUI’s tutorial gives no minimum version and says to use the latest. If LTX nodes are missing, it suggests the nightly build.

  2. Get access. Open Lightricks/LTX-2.5 on Hugging Face, sign in and accept the licence. Do the same for the IC-LoRA Pixel Spatial Upscaler repository if you want DFR. Downloads fail without it.

  3. Load a template. Click Templates, search for LTX-2.5 and choose one from the table above.

  4. Download the models. The Workflow Overview panel lists missing files and fetches them. On an offline machine, place them by hand:

    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
    ├── text_encoders/gemma4_e2b_it_int8_convrot.safetensors   (optional prompt enhancer)
    ├── vae/ltx-2.5-video-vae-bf16.safetensors   (template default; -conv-bf16 is the lighter option)
    ├── vae/ltx-2.5-audio-vae-bf16.safetensors
    └── latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors
  5. Check Prompt Enhance. Lightricks’ guide says the template turns the enhancer on by default; ComfyUI’s tutorial says it is off by default and adds about one to two minutes when enabled. Look at the toggle on the node and decide.

  6. Set size and length. Lightricks documents the text-to-video defaults as 768 × 512, 97 frames and 24 fps, which the second stage doubles to 1536 × 1024. Width and height must divide by 32, and the frame count must be a multiple of 8 plus 1.

  7. Run. Stage 1 samples video and audio latents together with euler_ancestral at CFG 1 on a fixed sigma schedule, the spatial upscaler doubles the video latent, and stage 2 refines it. The template applies a short negative prompt itself.

You need Lightricks’ ComfyUI-LTXVideo custom node pack only for the workflows that are not in the template browser. Its example_workflows/2.5/ folder holds audio-to-video, text-to-audio, IC-LoRA control, inpainting, outpainting, HDR and Tiled Fusion graphs. Install it from ComfyUI Manager by searching for “LTXVideo”.

The official templates have no standard-LoRA node. Lightricks’ guide says to insert ComfyUI’s LoraLoaderModelOnly after the model output and route it to both sampler stages.

Outside ComfyUI, the same weights run in Lightricks’ ltx-pipelines package, in Diffusers (the 0.40.0 release notes list LTX2.5 pipelines, though the model card still says to install from the main branch), in WanGP and in stable-diffusion.cpp. If you came here from another model, our MiniMax H3 ComfyUI guide follows the same layout for that model.

LTX 2.5 GGUF builds

Lightricks publishes no GGUF files. The most complete set is unsloth/LTX-2.5-GGUF, created on 13 October 2026, with both the distilled and the dev transformer. Unsloth measured each file against the BF16 original on the same prompts and seeds at 1280 × 704 and 121 frames. LPIPS is a perceptual distance, so lower is better.

Distilled fileSizeBits per weightLPIPS vs BF16
Q2_K8.46 GB3.220.567
UD-Q2_K_XL9.50 GB3.620.416
Q3_K_M11.30 GB4.300.390
UD-Q3_K_XL12.00 GB4.570.310
Q4_K_M13.58 GB5.170.352
UD-Q4_K_XL14.00 GB5.330.320
Q5_K_M15.63 GB5.950.316
UD-Q5_K_XL16.00 GB6.090.296
Q6_K17.81 GB6.780.274
Q8_022.78 GB8.680.242

These are Unsloth’s own measurements, not independent ones. Unsloth notes that the samplers re-inject noise at every step, so small weight errors change the motion more than the look of a frame, and that every file kept the subject, scene and prompt. Its recommendation is UD-Q4_K_XL for most GPUs, Q8_0 for output closest to BF16, and UD-Q2_K_XL on small cards. The UD- files are mixed precision: each layer’s type is chosen from a sensitivity scan. Our explainer on GGUF quantisation levels decodes the naming.

To use one in ComfyUI, Unsloth’s card installs the ComfyUI-GGUF custom node and links the files into place:

cd ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI-GGUF.git
pip install -r ComfyUI-GGUF/requirements.txt
cd ../models
ln -s "$(hf download unsloth/LTX-2.5-GGUF distilled/ltx-2.5-22b-distilled-UD-Q4_K_XL.gguf --quiet)" unet/.

Then open the built-in text-to-video template and replace its Load Diffusion Model node with Unet Loader (GGUF). The text encoder, VAEs and upscaler stay as Lightricks’ files. Unsloth runs the distilled files with a two-stage 8 + 3 step recipe and the dev files at 30 steps and CFG 3.

Other community repositories exist, among them Abiray’s and vantagewithai’s distilled GGUFs and a GGUF conversion of the text encoder. Sizes differ between publishers for the same label, so compare file sizes, not names. stable-diffusion.cpp added LTX-2.5 on 30 August 2026; its documentation requires the conv video VAE, because the diffusion decoder is not implemented there.

GGUF files are derivatives under the licence, so its terms apply to them too.

LTX 2.5 vs 2.3: what changed

LTX 2.3LTX 2.5
On Hugging Face sinceMarch 2026August 2026
Transformer22B, dev and distilled22B, dev and distilled
PackagingSingle 46.15 GB checkpoint per variantSplit pack, one file per component
Text encoderGemma 3, downloaded separatelyCustom Gemma 4 12B with projection bundled
Video decoderConvolutional VAEDiffusion decoder, with a conv VAE as the lighter option
Shots per generationOne continuous shotNative multishot
Official quantised filesSeparate FP8 and NVFP4 repositoriesINT8 ConvRot and NVFP4 in the main repository
Extras×1.5 and ×2 spatial upscalersDFR pipeline, prompt enhancer, duration head, EXR and HDR paths
LicenceLTX-2 Community LicenseLTX-2.x Community License

Assets do not carry over cleanly. Lightricks’ compatibility page says to keep 2.3 and 2.5 workflows and files separate unless a workflow or model card states otherwise. The model card adds that most LoRAs trained on 2.3 reportedly work on 2.5 unchanged but a few do not, so validate each adapter.

LTX 2.3 is not retired. Lightricks’ documentation still routes video-to-audio Foley, Dub-It lip-synced dialogue replacement and Relight to LTX 2.3 workflows, and warns against substituting 2.5 assets there.

LTX 2.5 prompt guide: the official rules

Lightricks’ prompting guide is specific, and it differs from the keyword lists that suit image models.

Single shot. Write one flowing paragraph in the present tense, about four to eight sentences. Cover six things: the shot, the scene (lighting, palette, texture), the action as a sequence with active verbs, the characters with physical cues instead of emotion labels, the camera movement, and the audio. Close-ups need more detail than wide shots.

Audio. Sound is generated with the picture, so describe it. Put dialogue in quotation marks and state language and accent when they matter. Name ambient sound and music.

Multishot. Keep the whole scene in one chronological paragraph and prefer two to four shots. At each cut:

  1. Name the transition in plain words: a hard cut, a match cut, a dissolve.
  2. Re-establish the new shot: scale, angle, who is in frame.
  3. Reuse the same identifiers for recurring subjects.
  4. Say whether music, dialogue or ambience continues.

Avoid numbered shot lists and screenplay sluglines unless the cut is also described in prose.

Image to video. Describe what happens next, not what the image already shows, and prefer a single continuous take.

Pacing. The duration predictor times the clip to the action as written and will not add pauses you did not ask for. Write “she pauses” or “a beat of silence” into the prompt, or set the length yourself.

What to avoid. Crowded frames, mixed light sources, chaotic motion, exact on-screen text beyond a short prominent word, and prompts pasted from other video models. The guide suggests running those through the prompt enhancer instead. It says nothing on negative prompts.

Resolution, duration and audio

Locally, the limits are arithmetic and memory, not a fixed menu. Dimensions must divide by 32, so UHD is 3840 × 2176, not 2160. Frame counts follow 8n + 1: 97 frames is about 4 seconds at 24 fps and 121 frames about 5.

The DFR template is the route to higher output. ComfyUI’s tutorial describes three stages on the distilled transformer: 8 steps at half the output size, which also generate one keyframe per second; 3 steps at full size with the detailing IC-LoRA; and an optional 4-step temporal pass that doubles a 5-second clip to 241 frames at 48 fps. Audio comes from stage 1 and passes through unchanged. Lightricks’ DFR command defaults to 1024 × 1536 at 24 fps and reaches 4K at 3840 × 2176, where its changelog names --spatial-upscalings 2 as the higher-quality route.

For comparison, the hosted ltx-2-5-fast model offers 6 to 20 seconds at 720p and 1080p when the frame rate is 24 or 25 fps, and 6 to 10 seconds at 1440p and 4K. Lightricks publishes no equivalent ceiling for the open weights.

How fast is LTX 2.5 locally?

Lightricks publishes one figure: its product page says that, self-hosted on two GB200 GPUs at 720p, LTX-2.5 generates a 10-second clip in 6.8 seconds, which it calls faster than real time. That is Lightricks’ own measurement on data-centre hardware, not a guide to a workstation card. Its quick-start says only that generation time depends on your GPU and VRAM, and neither ComfyUI’s tutorials nor Unsloth’s card give seconds per clip. Measure on your own card and record resolution, frame count, step counts and whether the prompt enhancer ran. For still images, see our round-up of image models for local hardware.

LTX 2.5 licence: the $10 million line

The weights use the LTX-2.x Community License Agreement, dated 11 August 2026 and amended on 29 October 2026. The code in the LTX-2 repository is Apache 2.0, which grants no rights in the model. The main terms:

  • Revenue threshold. Entities with annual revenue of at least $10,000,000 must buy a Commercial Use Agreement for any use that is not non-commercial. The model card says revenue is measured across the whole entity, including subsidiaries and affiliates under common control.
  • Non-commercial carve-out. Such an entity may still test, evaluate and do non-commercial research in a non-production environment. Revenue-generating use, anything that touches end users, and training, fine-tuning or distilling a model for commercial use all fall outside it.
  • Derivatives. Fine-tunes, LoRAs and quantised files are derivatives. Passing one to a company above the threshold does not license that company; it needs its own agreement.
  • Outputs and provenance. Lightricks claims no rights in what you generate. You may not remove or defeat watermarking, metadata or provenance features, and you are responsible for AI rules that apply to you, with the EU AI Act named in the text.

Our open-weight licence review checklist lists the questions to take to counsel. Nothing here is legal advice.

How VDF AI fits

VDF AI does not generate video, and it does not run, host, resell or integrate with LTX 2.5, ComfyUI or any of the community builds named here.

What a single workstation set-up leaves open is the organisational side: who approved the model and its licence, who may use it, and what record exists afterwards. VDF AI addresses those questions for language models and agents, on premises or air-gapped, with governed agents, role-based access control and audit logs. VDF AI Router registers Ollama and custom on-premises deployments beside the cloud models your policy allows, and its air-gap mode keeps routing to local models only.

Sources

Lightricks, ComfyUI and community pages for LTX 2.5, each opened on 14 November 2026.


Running open models on hardware you control? See how VDF AI Router keeps language-model traffic on approved local endpoints, or talk to us.

Frequently asked questions

How much VRAM does LTX 2.5 need?

Lightricks' system requirements state an NVIDIA GPU with at least 32 GB of VRAM, 32 GB of system RAM and 100 GB of disk, and recommend an 80 GB A100 or an H100. The files allow less: the official INT8 ConvRot transformer is 21.50 GB, the NVFP4 build 18.72 GB, and Unsloth's GGUF builds run from 8.46 GB at Q2_K to 22.78 GB at Q8_0. By our estimate from those sizes, 24 GB cards suit INT8 and 16 GB cards suit the 14.0 GB UD-Q4_K_XL file, both below the stated minimum.

Is there an LTX 2.5 GGUF?

Yes, from the community rather than from Lightricks. Unsloth's LTX-2.5-GGUF repository, created on 13 October 2026, holds the distilled and dev transformers from Q2_K at 8.46 GB to Q8_0 at 22.78 GB, plus mixed-precision UD files. Unsloth recommends UD-Q4_K_XL at 14.0 GB for most GPUs. In ComfyUI the files load through the ComfyUI-GGUF custom node with its Unet Loader (GGUF) node, and stable-diffusion.cpp has supported LTX-2.5 since 30 August 2026. The text encoder and VAEs still come from Lightricks' repository.

How do I run LTX 2.5 in ComfyUI?

Update ComfyUI to 0.32.0 or later, accept the licence on the gated Lightricks/LTX-2.5 repository on Hugging Face, then open the template browser, search for LTX-2.5 and pick Text to Video, Image to Video or First-Frame / Last-Frame. The Workflow Overview panel downloads missing models. The text-to-video template starts at 768 by 512 pixels and 97 frames at 24 fps, then upscales to 1536 by 1024. Advanced workflows such as audio-to-video and IC-LoRA control need Lightricks' ComfyUI-LTXVideo node pack.

What is the difference between LTX 2.5 and LTX 2.3?

Both are 22B models that generate video and audio together, but LTX 2.5 changes most of the parts around the transformer. It adds multishot generation, a diffusion video decoder, a custom Gemma 4 12B text encoder in place of Gemma 3, an optional prompt enhancer, automatic duration and the DFR pipeline. It ships as separate files instead of one 46.15 GB checkpoint, with official INT8 and NVFP4 builds, and it uses the newer LTX-2.x Community License. Foley, Dub-It and Relight workflows still use LTX 2.3 assets.

Is LTX 2.5 free for commercial use?

For organisations with less than 10 million US dollars in annual revenue, yes: the LTX-2.x Community License, dated 11 August 2026 and amended on 29 October 2026, allows commercial and production use at no cost. Entities at or above that figure, counted with affiliates under common control, need a paid Commercial Use Agreement, although they may test and evaluate the model in non-production environments. Fine-tuning or distilling for commercial use counts as commercial. This is general information, not legal advice.

How should I write prompts for LTX 2.5?

Lightricks' prompting guide asks for one flowing paragraph in the present tense, about four to eight sentences for a single shot, covering the shot, the scene, the action, the characters, the camera movement and the audio. Put spoken lines in quotation marks and name the language or accent if it matters. For multishot clips, keep to two to four shots, name each cut in plain words such as a hard cut or a dissolve, and re-establish who is in frame. The guide does not cover negative prompts.

Filed under
LTX-2.5video generationComfyUIGGUFopen-weight modelslocal AI hardware
On-Prem AI

Plan your on-prem AI deployment

Book an architecture call and we will scope a private, on-prem AI deployment for your environment — integrations, hardware, and governance included.

Keep reading