LTX-2.5-Comfy

About

What LTX-2.5-Comfy is

LTX-2.5 22B — Comfy - GGUF GGUF quantizations of Lightricks' LTX-2.5, converted for efficient execution in tools like ComfyUI. Example video is rendered using Q4KM quantize variant. Model Explanation LTX-2.5 is an open-world model with open weights, built for local execution and fine-tuning. It natively generates synchronized, high-fidelity video and audio in a single pass from text, image, and video inputs. Key Features of LTX-2.5: - Synchronized Video and Audio: A dual-stream DiT with a 4096-wide video path, a 2048-wide audio path, and cross-modal attention joining them. No separate audio stage required. - Native Multishot Generation: Generates connected scenes in a single pass: multiple shots that hold character identity, environment, lighting, voice, and visual style across cuts (whereas previous versions produced a single continuous shot). - Diffusion Fidelity Rendering: Dynamically allocates compute by scene complexity and budget, rendering flawless detail where it matters and remaining efficient everywhere else. - Custom Gemma 4 12B Text Encoder: Holds complex prompts together (multiple characters, camera moves, lighting, actions) instead of dropping details across a longer sequence. The original bf16 transformer weighs in at 39 GB. These GGUF quantizations bring it down to a manageable 11-23 GB for consumer hardware. GGUF Quantization VRAM Requirements Below is a reference guide for the approximate VRAM required to load the 22 Billion parameter LTX-2.5-Distilled model at various GGUF quantization levels. Note: The sizes below refer specifically to the raw model weights loaded into memory. Generating video or high-resolution images will require an additional 4-8 GB of VRAM for the execution context, VAE, and text encoders depending on your ComfyUI configuration. Performance vs Quality - Q4KM: Best balance for typical consumer hardware (like a 16GB RTX 4080). Minor quality loss but runs very fast. - Q5KM: The sweet spot if you have a 24GB card (RTX 3090 / 4090). Noticeably better coherence than Q4 without hitting memory limits. - Q6K: Very close to uncompressed BF16. Excellent choice if you have a powerful setup or Mac Studio. - Q80: Visually indistinguish

Summarised from the published model card. Read the full card on the HuggingFace links below.

Specifications

What it is

How it works

How video models work

Variants

Sizes & precisions

Open weights ship in several sizes and precisions. One page, all the variants — pick the one that fits your GPU. VRAM figures are estimates from model size.

Using it via the API

Call it like any OpenAI endpoint

Once AxForge deploys ltx-2-5-comfy for you, it answers on the OpenAI-compatible API — the same base URL and keys as every other model. (ltx-2-5-comfy below is illustrative; you get the exact model name on deployment.)

$ curl -sS https://api.axforge.ai/v1/videos/generations \
  -H "Authorization: Bearer $AXFORGE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"ltx-2-5-comfy","prompt":"a drone shot over a forest"}'

Details

Languages, data & research

Licence

Commercial licence needed

Sources

Weights & code

Explore

More video models