Where do I download LTX-2.5 models?

From the official Hugging Face repository Lightricks/LTX-2.5 (https://huggingface.co/Lightricks/LTX-2.5). Every row in the tables above links straight to the file on Hugging Face; the IC-LoRAs live in their own Lightricks repositories. Inside ComfyUI, Workflow Overview → Download all fetches the same files automatically.

How much disk space do I need?

The minimum set for the two-stage T2V/I2V graph is about 72.7 GB in bf16 (42 GB transformer, 26.3 GB text encoder, plus VAEs and upscaler). Swapping in the INT8 transformer and INT8 text encoder brings that down to about 41.3 GB. Adding the full dev model, temporal upscaler, distilled LoRA and IC-LoRAs adds to the total.

INT8 vs NVFP4 vs bf16 — which transformer should I use?

bf16 (42 GB) is the reference build for 32 GB+ cards. INT8 (21.5 GB) is about half the size and the usual choice for 24 GB cards. NVFP4 (18.7 GB) is the smallest but requires an NVIDIA GPU with FP4 support (Blackwell, e.g. RTX 50-series). All three load with the same workflow — only the file in the loader node changes.

Do I need the Gemma text encoder?

Yes. Every LTX-2.5 graph — video, audio and control alike — runs your prompt through the Gemma 4 12B text encoder with the LTX-2.5 projection. Use the 26.3 GB bf16 file on 48 GB+ cards, or the 15.4 GB INT8 file everywhere else.

Can I try LTX-2 without downloading anything?

Yes. The online generator on this site runs the same LTX-2 model on hosted GPUs: text-to-video and image-to-video, up to 4K with synchronized audio, and new accounts get free credits. Nothing to install and no VRAM limit on your side.

Official file list for ComfyUI-LTXVideo 2.5

LTX-2.5 Model Files for ComfyUI

Every file the official LTX-2.5 example workflows reference — where it lives on Hugging Face, which ComfyUI/models/ subfolder it belongs in, how big it is, and which lower-VRAM variants exist.

Weights are © Lightricks and distributed under the LTX-2 Community License. All links go to the official Hugging Face repositories — we do not mirror weights.

Browse workflows Skip the 100 GB download — generate online

72.7 GB

minimum set, bf16

41.3 GB

minimum set with INT8 variants

6 folders

under ComfyUI/models/

32 GB+ VRAM

Lightricks' official requirement

Quick start (minimum set)

The six files needed for the recommended Text / Image to Video two-stage graph. Everything else on this page is optional.

ltx-2.5-22b-distilled-transformer-bf16.safetensors
ComfyUI/models/diffusion_models/
42.0 GBHugging Face
gemma4-12b-with-proj-ltx-2.5-bf16.safetensors
ComfyUI/models/text_encoders/
26.3 GBHugging Face
ltx-2.5-video-vae-bf16.safetensors
ComfyUI/models/vae/
1.5 GBHugging Face
ltx-2.5-video-vae-conv-bf16.safetensors
ComfyUI/models/vae/
1.5 GBHugging Face
ltx-2.5-audio-vae-bf16.safetensors
ComfyUI/models/vae/
0.40 GBHugging Face
ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors
ComfyUI/models/latent_upscale_models/
1.0 GBHugging Face

Total download72.7 GB

You do not have to fetch these by hand: open the workflow in ComfyUI, click Workflow Overview → Download all, and ComfyUI places each file in the right subfolder.

All LTX-2.5 files

Grouped by the ComfyUI/models/ subfolder they go in. Lower-VRAM variants sit under the model they replace.

Diffusion models

File	Folder	Size	Purpose	Link
`ltx-2.5-22b-distilled-transformer-bf16.safetensors`	`ComfyUI/models/diffusion_models/`	42.0 GB	The LTX-2.5 22B distilled video model — used by every example workflow. Few sampling steps, built for fast iteration.	Hugging Face
↳`ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors`lower-VRAM variant	`ComfyUI/models/diffusion_models/`	21.5 GB	INT8 — about half the size; the usual pick for 24 GB cards.	Hugging Face
↳`ltx-2.5-22b-distilled-transformer-nvfp4.safetensors`lower-VRAM variant	`ComfyUI/models/diffusion_models/`	18.7 GB	NVFP4 — smallest; needs an NVIDIA GPU with FP4 support (Blackwell).	Hugging Face
`ltx-2.5-22b-dev-transformer-bf16.safetensors`	`ComfyUI/models/diffusion_models/`	42.0 GB	Full (non-distilled) LTX-2.5 model. Higher quality ceiling with more sampling steps; not used by the distilled example workflows.	Hugging Face
↳`ltx-2.5-22b-dev-transformer-comfy-int8-convrot.safetensors`lower-VRAM variant	`ComfyUI/models/diffusion_models/`	21.5 GB	INT8 variant of the full model.	Hugging Face

Text encoders

File	Folder	Size	Purpose	Link
`gemma4-12b-with-proj-ltx-2.5-bf16.safetensors`	`ComfyUI/models/text_encoders/`	26.3 GB	Gemma 4 12B text encoder with the LTX-2.5 projection — turns your prompt into conditioning. Required by all workflows.	Hugging Face
↳`gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors`lower-VRAM variant	`ComfyUI/models/text_encoders/`	15.4 GB	INT8 text encoder — saves ~11 GB on disk and in memory.	Hugging Face

VAEs

File	Folder	Size	Purpose	Link
`ltx-2.5-video-vae-bf16.safetensors`	`ComfyUI/models/vae/`	1.5 GB	Video VAE — encodes input frames and decodes the generated latents to pixels.	Hugging Face
`ltx-2.5-video-vae-conv-bf16.safetensors`	`ComfyUI/models/vae/`	1.5 GB	Convolutional video VAE variant referenced by the example graphs' decode path.	Hugging Face
`ltx-2.5-audio-vae-bf16.safetensors`	`ComfyUI/models/vae/`	0.40 GB	Audio VAE — decodes the synchronized soundtrack LTX-2.5 generates alongside the video.	Hugging Face

Latent upscalers

File	Folder	Size	Purpose	Link
`ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors`	`ComfyUI/models/latent_upscale_models/`	1.0 GB	2× latent spatial upscaler — the second stage of every two-stage workflow.	Hugging Face
`ltx-2.5-latent-temporal-upscaler-x2-bf16-1.0.safetensors`	`ComfyUI/models/latent_upscale_models/`	0.30 GB	2× temporal upscaler — doubles frame rate in latent space. Optional; not wired in the default example graphs.	Hugging Face

LoRAs

File	Folder	Size	Purpose	Link
`ltx-2.5-22b-distilled-lora-450-bf16.safetensors`	`ComfyUI/models/loras/`	8.9 GB	Distillation LoRA for running the full dev model in few steps. Not needed when you use the distilled transformer.	Hugging Face
`ltx-2.3-22b-ic-lora-union-control-ref0.5.safetensors`	`ComfyUI/models/loras/`	—	IC-LoRA: depth / canny / pose control from a reference video.	Hugging Face
`ltx-2.3-22b-ic-lora-motion-track-control-ref0.5.safetensors`	`ComfyUI/models/loras/`	—	IC-LoRA: follow motion tracks drawn on the first frame.	Hugging Face
`ltx-2.3-22b-ic-lora-ingredients-0.9.safetensors`	`ComfyUI/models/loras/`	—	IC-LoRA: generate from a reference sheet of characters, props and locations.	Hugging Face
`ltx-2.3-22b-ic-lora-in-outpainting-0.9.safetensors`	`ComfyUI/models/loras/`	—	IC-LoRA: inpaint a masked region or outpaint beyond the frame of a reference video.	Hugging Face
`ltx-2.3-22b-ic-lora-deblur-0.9.safetensors`	`ComfyUI/models/loras/`	—	IC-LoRA example shipped with the V2V graph (deblur). Swap for any identity / style / edit LoRA.	Hugging Face

The IC-LoRAs are LTX-2.3-era weights: the official 2.5 control graphs apply them on top of the 2.5 distilled backbone, which is exactly how Lightricks ships them.

Model patches

File	Folder	Size	Purpose	Link
`ltx-2.5-duration-head-bf16.safetensors`	`ComfyUI/models/model_patches/`	0.05 GB	Small duration-prediction head used by audio workflows.	Hugging Face

Which variant for my GPU?

Pick the transformer and text-encoder build that fits your card, then download only those.

VRAM	Example cards	Transformer	Text encoder	Verdict	Advice
Under 16 GB	RTX 3060 12G, 4060, 4070, most laptops	`nvfp4`	`int8`	Use online	The 22B transformer alone is 18.7 GB even at NVFP4. Expect out-of-memory errors or extreme offloading. Generate online instead — same LTX-2 model, no setup.
16 GB	RTX 4060 Ti 16G, 4080, 5070 Ti, A4000	`nvfp4`	`int8`	Tight	Borderline. Use the NVFP4 transformer (Blackwell) or INT8 with aggressive offloading (--reserve-vram), the INT8 text encoder, single-stage graphs and low resolution. Slow, and not every graph will fit.
24 GB	RTX 3090, 4090, A5000	`int8`	`int8`	Workable	Workable with the INT8 transformer + INT8 text encoder and the low-VRAM loader nodes. Two-stage graphs run; keep decode tiles small. Leave bf16 files for 32 GB+ cards.
32 GB	RTX 5090, V100 32G	`bf16`	`int8`	Comfortable	Lightricks' stated minimum. bf16 transformer with the low-VRAM loaders fits; INT8 text encoder frees headroom for larger decode tiles and longer clips.
48 GB+	RTX 6000 Ada, A6000, H100	`bf16`	`bf16`	Comfortable	Everything in bf16 without offloading. Increase the decode tile size (see the Decode notes in each graph) for faster runs.

Lightricks' official requirement is 32 GB+ VRAM. The lower tiers above are conservative community guidance, not an official support matrix.

Under 24 GB? Generate with the same LTX-2 model online

4K output with synchronized audio, nothing to download, free credits to start.

Open the online generator

Folder layout

Match these subfolder names exactly — ComfyUI only lists files from the folder that matches the loader node.

ComfyUI/
└── models/
    ├── diffusion_models/
    │   └── ltx-2.5-22b-distilled-transformer-bf16.safetensors
    ├── text_encoders/
    │   └── gemma4-12b-with-proj-ltx-2.5-bf16.safetensors
    ├── vae/
    │   └── ltx-2.5-video-vae-bf16.safetensors
    ├── latent_upscale_models/
    │   └── ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors
    ├── loras/
    │   └── ltx-2.3-22b-ic-lora-union-control-ref0.5.safetensors
    └── model_patches/
        └── ltx-2.5-duration-head-bf16.safetensors

Legacy LTX-2.3 files

Only needed for the LTX-2.3 workflows listed on the workflows page. New setups should start with 2.5. See the 2.3 workflows

File	Folder	Size	Purpose	Link
`ltx-2.3-22b-distilled-1.1.safetensors`	`ComfyUI/models/checkpoints/`	—	LTX-2.3 distilled checkpoint (all-in-one).	Hugging Face
`ltx-2.3-22b-dev.safetensors`	`ComfyUI/models/checkpoints/`	—	LTX-2.3 full checkpoint.	Hugging Face
`ltx-2.3-spatial-upscaler-x2-1.1.safetensors`	`ComfyUI/models/latent_upscale_models/`	—	2× spatial upscaler for 2.3 two-stage graphs.	Hugging Face
`ltx-2.3-temporal-upscaler-x2-1.0.safetensors`	`ComfyUI/models/latent_upscale_models/`	—	2× temporal upscaler for 2.3.	Hugging Face
`ltx-2.3-22b-distilled-lora-384-1.1.safetensors`	`ComfyUI/models/loras/`	—	Distillation LoRA for the 2.3 dev checkpoint.	Hugging Face
`gemma-3-12b-it-qat-q4_0-unquantized (whole repo)`	`ComfyUI/models/text_encoders/gemma-3-12b-it-qat-q4_0-unquantized/`	—	Gemma 3 12B text encoder used by LTX-2.3 graphs.	Hugging Face

LTX-2.5 Model Files for ComfyUI

Quick start (minimum set)

All LTX-2.5 files

Diffusion models

Text encoders

VAEs

Latent upscalers

LoRAs

Model patches

Which variant for my GPU?

Under 24 GB? Generate with the same LTX-2 model online

Folder layout

Legacy LTX-2.3 files

Common download problems

LTX-2.5 model FAQ

Keep exploring

ComfyUI workflows

What is LTX-2.5?

How to use LTX-2

LTX-2.5 Model Files for ComfyUI

Quick start (minimum set)

All LTX-2.5 files

Diffusion models

Text encoders

VAEs

Latent upscalers

LoRAs

Model patches

Which variant for my GPU?

Under 24 GB? Generate with the same LTX-2 model online

Folder layout

Legacy LTX-2.3 files

Common download problems

"Missing models" when opening the workflow

CUDA out of memory

Missing custom nodes / red nodes

LTX-2.5 model FAQ

Where do I download LTX-2.5 models?

How much disk space do I need?

INT8 vs NVFP4 vs bf16 — which transformer should I use?

Do I need the Gemma text encoder?

Can I try LTX-2 without downloading anything?

Keep exploring

ComfyUI workflows

What is LTX-2.5?

How to use LTX-2