Model overview · Lightricks
LTX 2.5
LTX-2.5 is Lightricks’ open-weights world model for local execution and fine-tuning. Its established use is generating synchronized high-fidelity video and audio from text, image, and video inputs; robotics / physical AI applicability is still developing. Hugging Face: Lightricks/LTX-2.5 (Diffusers pack: Lightricks/LTX-2.5-Diffusers).
What it is
LTX-2.5 ships as a split, Comfy-aligned pack (one .safetensors per component), not a monolith. Core pieces:
- DiT transformers: distilled / full (dev) bf16, plus Comfy int8 and NVFP4 variants.
- Text encoder: custom Gemma 4 12B + projections.
- VAEs: new diffusion video decoder (DiffVAE) and faster Conv VAE; audio VAE + vocoder.
- Optional duration head, spatial/temporal latent upscalers, distilled LoRA, etc.
Constraints from the card: num_frames % 8 == 1; width/height divisible by 32.
Who made it
Built by Lightricks. Weights: Lightricks/LTX-2.5; code: Lightricks/LTX-2. License: LTX-2.x Community License (HF license: other). The card states commercial/production use at no cost under that license below a revenue threshold; transfer of fine-tunes may need a paid license—see LICENSE. A paid Commercial Use Agreement is also offered.
What it is for
Established uses from the model card:
- Text-to-video + audio with synchronized sound.
- Image-to-video with first-frame / multi-frame conditioning.
- Native multishot: connected scenes in one pass with more consistent identity, lighting, voice, and style.
- Self-host via ComfyUI, official
ltx-pipelines, or Diffusers—no forced per-generation API.
How to try
Option A — Python (ltx-pipelines): clone the repo, uv sync, hf download the distilled split pack, run python -m ltx_pipelines.distilled with paths to transformer / text-encoder / VAEs / duration-head / spatial upsampler. Python ≥ 3.12, CUDA ≥ 12.7, PyTorch ≈ 2.7 recommended.
git clone https://github.com/Lightricks/LTX-2.git
cd LTX-2 && uv sync && source .venv/bin/activate
# then hf download weights per the model card and run the distilled moduleOption B — ComfyUI: official LTX-2.5 workflow templates.
Option C — Diffusers: Lightricks/LTX-2.5-Diffusers from Diffusers main; two-stage I2V example with latent upsample. An online API playground is also linked on the card.
Full commands: model card.
What is new in 2.5
From the model card’s “What’s new”:
- Native multishot generation.
- Diffusion fidelity rendering that allocates compute by scene complexity.
- New diffusion video decoder for sharper faces, textures, on-screen text, and motion.
- Custom Gemma 4 12B text encoder for complex prompts.
- Prompt enhancer and optional duration predictor.
- Improved distilled model retaining more full-model quality at speed.
Most LTX-2.3 LoRAs / IC-LoRAs run on 2.5; validate before production.
FAQ
Distilled vs dev?
Distilled uses a fixed ~8-step schedule with CFG=1 (faster). Dev is the full trainable DiT. Comfy int8 files are ComfyUI-only—not for ltx-pipelines / plain PyTorch.
License basics?
LTX-2.x Community License: card says free commercial use under a revenue threshold; fine-tune transfer may need a paid license. Read the repo LICENSE.
Low VRAM tips?
Card suggests --quantization fp8-cast and --offload cpu on bf16 transformers, or Comfy / NVFP4 paths with the right hardware.
Frame constraints?
num_frames % 8 == 1; width/height divisible by 32. Omit --num-frames to let the duration head choose from the prompt.
Is this official?
No. Unofficial overview. Lightricks HF + GitHub are authoritative.
Unofficial model overview. Summarizes the public Hugging Face model card for LTX-2.5. Not affiliated with Lightricks. Source of truth: Lightricks/LTX-2.5.