LTX-2.5 in ComfyUI: The Upgrade Guide

This is the LTX-2.5 upgrade guide. If you already run LTX-2.3 in ComfyUI, on a CLI, or with a LoRA you trained yourself, we’ll answer the three questions you probably have: what changed, what you have to redo, and what is worth trying first. We walk the same workflow you already know, swap in the 2.5 checkpoint and the Gemma 4 encoder, and show what happens to prompts you already wrote. The walkthrough covers native multi-shot prompting and the two rules that make it work, carrying the action through the cut and naming your subject the same way in every shot. Then a same seed and same sampler comparison against 2.3 to show where it stops holding a coherent shot list, the Gemma 3 to Gemma 4 text encoder and what happens when you ask for five things at once, the prompt enhancer turning one sentence into 760 characters and the negative prompt it never touches, the two decoders 2.5 ships with and the kind of footage where the difference actually shows, and Diffusion Fidelity Rendering, which plants a full-detail keyframe about once a second and drops them before export. On the migration question, no retraining. Your 2.3 LoRAs load against a 2.5 checkpoint at a default strength of 1.2. We close on Hugging Face gating and the 403 that catches most people, the new split pack, and why bf16 and int8 setups should never be mixed. This is for ComfyUI users, filmmakers, and technical creators with an LTX-2.3 workflow already running who want to move to 2.5 without rebuilding from scratch. Get started: GitHub: https://github.com/Lightricks/ComfyUI... HuggingFace: https://huggingface.co/Lightricks/LTX... Documentation: https://docs.ltx.video 00:00 - 00:29 Intro: One Prompt, One Generation, Not an Edit 00:29 - 00:56 The Three Questions This Upgrade Answers 00:56 - 01:17 Native Multi-shot: There Is No Multi-shot Flag 01:17 - 02:04 Carry the Action Through the Cut, Name Your Subject Every Time 02:04 - 02:35 Same Prompt Skeleton, Completely Different World 02:35 - 03:16 Same Seed, Same Sampler: LTX-2.3 vs LTX-2.5 03:16 - 04:00 Gemma 3 to Gemma 4: All Five Asks Survived 04:00 - 04:27 The Prompt Enhancer and the Negative Prompt It Never Touches 04:27 - 05:25 Two Decoders: Diffusion, Convolutional, and Where You See It 05:25 - 05:59 Do You Have to Retrain? Your 2.3 LoRAs Load on 2.5 05:59 - 06:54 Auto Duration and Diffusion Fidelity Rendering 06:54 - 07:40 Why One Latent Token Covers Eight Frames 07:40 - 08:28 Hidden Keyframes: The Storyboard the Model Builds for Itself 08:28 - 08:48 Standard Path vs DFR: Watch It Move 08:48 - 09:12 Gated Repos, 403 Errors, and IC-LoRA Access 09:12 - 09:37 The Split Pack and Why Your Old Download Commands Break 09:37 - 10:06 bf16 or int8: Neither Is Wrong, Do Not Mix Them 10:06 - 10:18 Final Thoughts 🔷 Join our Discord Community:   / discord   🔷 Follow us on X: https://x.com/ltx_io 🔷 Follow us on Instagram:   / ltx.io   About LTX LTX builds open world models: AI that understands and generates physical reality across video, audio, and multimodal inputs. Open, efficient, and yours to own. Run on your own hardware and fine-tune on your own IP, with 18M+ downloads and no vendor lock-in. 🔔 Subscribe for more LTX model releases, open-source tools, training guides, and developer resources.