LTX-2.5 Distilled GGUF
This repository provides quantized GGUF formats of the distilled transformer from Lightricks/LTX-2.5. These weights are highly optimized for local execution, allowing you to run high-fidelity video and audio generation workflows on hardware with memory constraints while retaining the core visual fidelity of the original base model.
Available Quantizations
| File | Size | Description |
|---|---|---|
| LTX-2.5-Distilled-Q3_K_S.gguf | 12.6 GB | Smallest footprint for extreme low-VRAM setups, but highest quality loss. |
| LTX-2.5-Distilled-Q3_K_M.gguf | 12.9 GB | Smallest footprint with the highest quantization loss. Best for strict memory limits. |
| LTX-2.5-Distilled-Q4_K_S.gguf | 15.3 GB | Slightly smaller than Q4_K_M, balancing speed and limited VRAM/RAM. |
| LTX-2.5-Distilled-Q4_K_M.gguf | 15.7 GB | Recommended baseline. Good balance of visual fidelity, motion consistency, and memory footprint. |
| LTX-2.5-Distilled-Q5_K_M.gguf | 18.1 GB | Higher precision, retaining strong prompt adherence with minimal degradation. |
| LTX-2.5-Distilled-Q6_K.gguf | 18.6 GB | Near-unquantized visual quality, very low quantization loss. |
| LTX-2.5-Distilled-Q8_0.gguf | 23.6 GB | Largest quantized footprint, nearly indistinguishable from the original bf16 weights. |
Required Supporting Components (Text Encoders & VAEs)
LTX-2.5 operates on a split-component architecture. To run these GGUF diffusion models in environments like ComfyUI or local Python pipelines, you must also fetch the official Text Encoders and VAEs directly from the upstream Lightricks/LTX-2.5 repository.
Text Encoders
Download these from the text_encoders directory:
gemma4-12b-with-proj-ltx-2.5-bf16.safetensors(26.3 GB) — The full unquantized Gemma 4 12B text encoder.gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors(15.4 GB) — An optimized INT8 text encoder specifically designed for lower VRAM usage in ComfyUI.
VAEs (Video & Audio)
Download these from the vae directory:
ltx-2.5-video-vae-bf16.safetensors(1.47 GB) — The primary DiffVAE for highest-quality video decoding.ltx-2.5-video-vae-conv-bf16.safetensors(1.45 GB) — Convolutional video VAE for faster, lighter processing.ltx-2.5-audio-vae-bf16.safetensors(365 MB) — Audio VAE and vocoder required for synchronized audio generation.
📈 Latent Upscalers (Recommended)
To achieve maximum visual fidelity and smoothness, it is highly recommended to use the official Stage 2 latent upscalers alongside this GGUF model.
You can download them from the official Lightricks LTX-2.5 Repository here.
Available Upscaler Models
- ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors (996 MB)
- Function: Doubles the pixel resolution of your generated video for sharp, 4K-like crispness.
- ltx-2.5-latent-temporal-upscaler-x2-bf16-1.0.safetensors (262 MB)
- Function: Doubles the framerate (e.g., 24fps to 48fps) by interpolating frames, resulting in buttery-smooth motion.
ComfyUI T2V and I2V GGUF Workflows
To make running this GGUF model as plug-and-play as possible, I have included two pre-configured ComfyUI workflows in this repository.
You can view and download them directly here:
- 🎥 Text-to-Video Workflow (T2V)
- Best for: Generating videos from scratch using only text prompts and audio conditioning.
- 🖼️ Image-to-Video Workflow (I2V)
- Best for: Animating a starting input image (like a portrait or landscape) while using a text prompt to guide the motion and camera tracking.
About the Original Model
LTX-2.5 is an open-world model built for local execution and fine-tuning. It specializes in generating synchronized, high-fidelity video and audio from text, image, and video inputs.
Key Features of LTX-2.5
- Native Multishot Generation: Generate connected scenes in a single pass holding character identity, environment, lighting, voice, and visual style across cuts.
- Diffusion Fidelity Rendering: The model dynamically allocates compute based on scene complexity, rendering flawless detail where needed.
- Distilled Efficiency: These GGUF checkpoints are derived from the distilled model, capturing much of the full 22B model's capabilities in a significantly faster, smaller package.
Usage Requirements
Ensure your inference engine (such as llama.cpp or compatible ComfyUI GGUF loader nodes) supports the LTX-2.5 DiT architecture. Depending on the quantization tier, appropriate system RAM or GPU VRAM size must be allocated to accommodate the file sizes listed above.
License & Limitations
These weights fall under the original LTX-2.x Community License. Commercial and production use is permitted at no cost for entities under $10M annual revenue.
- This model is not intended or able to provide factual information.
- Prompt following is heavily influenced by prompting style.
- The model may fail to generate videos that match the prompt perfectly or may generate artifacts in highly complex scenes.
For full architectural details, official multishot prompting guides, and citation information, please refer to the Original LTX-2.5 Model Card.