Home Creators Posts Import Register Favorites Logout

Content

Video : https://youtu.be/XAhuGRCI2tM

Hey everyone.

This video fixed of the video editor exporting problem from the last upload. https://youtu.be/XAhuGRCI2tM

New video is up, and this one covers LTX-2.5 — the latest open-weights video model from Lightricks. I've been testing it locally in ComfyUI, and there are some real changes under the hood compared to 2.3.

The headline feature is Diffusion Fidelity Rendering. Instead of applying a fixed compression rate to every frame like 2.3 did, LTX-2.5 reads each frame in real time and allocates compute where it's needed — faces, textures, motion get more power. Flat surfaces and simple backgrounds get less. Same total compute budget, just distributed by need. Think of it like adaptive bitrate in video streaming, but inside the diffusion process itself.

The other big addition is native multi-shot. You can write one prompt and get multiple connected shots — different camera angles, cut scenes — all holding character identity, environment, and lighting consistency across the cuts. No plugins or LoRAs needed.

In the video I run through the full local setup: downloading the model weights from Hugging Face, picking between BF16 and int8 convrot depending on your GPU, and walking through the official ComfyUI workflow from the LTX team. The two-stage pipeline takes you from 720p in stage one to 1080p or higher after the latent upscaler, and the whole thing runs in about 30 seconds on consumer hardware.

I also built an agent prompt skill for LTX-2.5 that handles the three prompt formats — single shot, multi-shot, and screenplay style. If you use AI agents to write your video prompts, it saves a lot of time.

One honest finding: for dialogue scenes and close-up character work, LTX-2.5 does a solid job. For high-action scenes with fast motion, it still needs work. The upscaling pipeline is solid. Not everything is perfect, and that's fine — the foundation is getting stronger.

— Benji

Try LTX-2.5 for yourself, LTX-2.5 Hugging Face Repo

https://huggingface.co/Lightricks/LTX-2.5

Open Weights LTX-2.5

https://ltx.io/2-5-open-weights

Official Site

https://ltx.io/model/ltx-2-5

ComfyUI-LTXVideo

https://github.com/Lightricks/ComfyUI-LTXVideo

Example Workflows: https://github.com/Lightricks/ComfyUI-LTXVideo/tree/master/example_workflows/2.5

LTX 2.5 Video Prompt Agent Skill

https://github.com/benjiyaya/LTX-2.5-Agent-Prompt-Skill

Files

LTX-2.5 Open Weights Running In ComfyUI With Multi-Shot & Diffusion Fidelity Rendering

A hands-on first look at LTX-2.5, the latest open-weights video generation model from LTX.io. Covers the new Diffusion Fidelity Rendering technology, native multi-shot generation, the custom Gemma 4 12B text encoder, and a full ComfyUI local setup walkthrough. Includes text-to-video, image-to-video with Krea 2, and two-stage 720p-to-1080 or higher upscaling. Try LTX-2.5 for yourself, Official model page: https://ltx.io/model/ltx-2-5 Open weights page: https://ltx.io/2-5-open-weights HuggingFace repo: https://huggingface.co/Lightricks/LTX-2.5 ComfyUI-LTXVideo: https://github.com/Lightricks/ComfyUI-LTXVideo Agent prompt skill: https://github.com/benjiyaya/LTX-2.5-Agent-Prompt-Skill Thank you ‪@LTX_io for supporting this video. Timeline: 0:00 - Introduction to LTX 2.5 and key changes from version 2.3. 0:18 - Explanation of Diffusion Fidelity Rendering and dynamic compute allocation. 1:04 - Technical specifications (22B parameters, 4K resolution, 16GB VRAM requirements). 1:24 - New features overview: Native multi-shot, updated Gemma 2 text encoder, and speed performance. 2:26 - How to download open weights from Hugging Face and file selection guide. 3:44 - ComfyUI installation and recommended custom nodes. 4:16 - Prompting strategies for LTX 2.5 (screenplay style and multi-shot formats). 4:55 - First look: Conversational scene tests and multi-shot consistency. 5:16 - Deep dive into the ComfyUI workflow (Model loaders, text encoders, and enhancers). 6:30 - Four-step generation process: Pre-processing, sampling stages, and upscaling. 7:34 - Detailed look at the new 2.5 latent spatial upscaler. 8:27 - Prompt engineering for multi-shot actions and courtroom drama examples. 9:02 - Video quality comparison: Text-to-video vs. Image-to-video tests. 11:12 - Demonstrating character-centric shots using Image-to-video. 12:56 - Final summary: Strengths in talking head shots and areas for improvement in action scenes. #LTX25 #AIVideo #OpenSourceAI #comfyui If You Like tutorial like this, You Can Support Our Work In Patreon: https://www.patreon.com/c/aifuturetech

Comments

No comments found for this post.