Sign In

LTX-2 19B GGUF — I2V + T2V, audio-aware, from 4GB VRAM with Triton, ComfyUI workflow

Download

2 variants available

Type
Workflows
Stats

9

Reviews

No reviews yet

Published

Nov 17, 2026

Base Model
Hash
AutoV2B9C73855F2

☕ One coffee = one happy creator: https://buymeacoffee.com/the_frizzy1


A simple colour-coded LTXV-2 workflow, image to video and text to video, audio-aware, on the 19B model as GGUF. Two files, one for I2V and one for T2V. It is light: it runs from 4 GB with Triton and it is comfortable at 12 GB.

What you need

✔ Custom nodes — ComfyUI-GGUF and KJNodes — KJNodes has to stay updated, v1.2.8 or newer
✔ ComfyUI version — Not stated in the README. Keep ComfyUI itself updated — stale nodes are what breaks this one.
✔ VRAM — 4 GB with Triton is the floor · 12 GB is comfortable
✔ System RAM — Not measured. The 3060 machine had 48 GB in it; no peak figure was taken.
✔ Tested on — RTX 3060 12 GB · 48 GB RAM · and RTX 3050 4 GB with Triton
✔ Licence — Workflow MIT · LTX-2 Community License Agreement · Gemma 3 text encoder: Gemma Terms of Use

Models to download

Click a file name to download it. Put it in the folder shown, inside ComfyUI/models/. File names are exactly what the workflow .json loads; subfolders inside the model folder are fine.

Required — 6 files. Three of them have no size in my downloads file, so there is no honest total to give.

diffusion_models/

File: ltx-2-19b-distilled_Q4_K_M.gguf — 12.7 GB
Download:
https://huggingface.co/Kijai/LTXV2_comfy/resolve/main/diffusion_models/ltx-2-19b-distilled_Q4_K_M.gguf

Put it here:
ComfyUI/models/diffusion_models/

The distilled model both .json files load. _Q8_0 is the bigger quant.

text_encoders/

File: gemma_3_12B_it_fp8_e4m3fn.safetensors — 13.2 GB
Download:
https://huggingface.co/GitMylo/LTX-2-comfy_gemma_fp8_e4m3fn/resolve/main/gemma_3_12B_it_fp8_e4m3fn.safetensors

Put it here:
ComfyUI/models/text_encoders/

Gemma 3 12B text encoder

File: ltx-2-19b-embeddings_connector_dev_bf16.safetensors — not listed
Download:
https://huggingface.co/Kijai/LTXV2_comfy/resolve/main/text_encoders/ltx-2-19b-embeddings_connector_dev_bf16.safetensors

Put it here:
ComfyUI/models/text_encoders/

the embeddings connector

vae/

File: LTX2_video_vae_bf16.safetensors — not listed
Download:
https://huggingface.co/Kijai/LTXV2_comfy/resolve/main/VAE/LTX2_video_vae_bf16.safetensors

Put it here:
ComfyUI/models/vae/

video VAE

File: LTX2_audio_vae_bf16.safetensors — not listed
Download:
https://huggingface.co/Kijai/LTXV2_comfy/resolve/main/VAE/LTX2_audio_vae_bf16.safetensors

Put it here:
ComfyUI/models/vae/

audio VAE

latent_upscale_models/

File: ltx-2-spatial-upscaler-x2-1.0.safetensors — 1.0 GB
Download:
https://huggingface.co/Lightricks/LTX-2/resolve/main/ltx-2-spatial-upscaler-x2-1.0.safetensors

Put it here:
ComfyUI/models/latent_upscale_models/

2x latent upscaler

Optional — the distill LoRA. Its loader ships bypassed.

loras/

File: ltx-2-19b-distilled-lora-384.safetensors — 7.7 GB
Download:
https://huggingface.co/Lightricks/LTX-2/resolve/main/ltx-2-19b-distilled-lora-384.safetensors

Put it here:
ComfyUI/models/loras/

No measured numbers for this one

I have no benchmark table for this workflow. No wall times, no peak VRAM, no resolution-by-resolution figures, on either card. The 4 GB floor is a "it ran" claim, not a measured run, and the RTX 3050 has no timings at all. I am not going to make numbers up to fill the gap — if you need to know what it costs on your card, run it once and watch it.

What breaks

  • LTXVEmptyLatentAudio – AttributeError means your KJNodes is too old. Update to v1.2.8 or newer.

  • Checkerboard noise in the output means update all your nodes and ComfyUI itself.

  • 4 GB is the floor and it needs Triton to get there. Without Triton that card is not going to make it.

  • There is a newer workflow for this family: LTX 2.5 Ultimate for 12 GB. Unless you are on a 4 GB card, look at that one first.

Get the workflow

More videos on this stuff: https://www.youtube.com/playlist?list=PLQ3PcJNbU2Ls