Today we are releasing LTX-2.5, the most capable open weights model for the media and entertainment industry. LTX-2.5 renders sharper, more physically accurate images and holds continuity across sequences rather than a single output, optimizing traditional workflows and unlocking new possibilities for the filmmaking community. LTX is open weights, so production teams can get started instantly on their own hardware today.
Quality That Holds Up On Screen
LTX-2.5 also introduces Diffusion Fidelity Rendering, a new rendering approach designed to push visual fidelity further. The model first builds the motion and structure of the video in an 8x temporally compressed latent space, while generating high-fidelity keyframes alongside it to anchor visual detail, dynamically allocating more compute to the visually demanding moments and less where it is not needed. Once the structure is established, a dedicated diffusion-based rendering stage brings that detail into the final video. The result is footage that holds its fidelity frame-to-frame, for premium-quality generation.
Building off of our belief in supporting high quality image outputs, LTX-2.5 augments the earlier Variational Autoencoder (VAE) decoder with a new diffusion video decoder that reconstructs fine detail like faces which had previously softened under compression. Outputs from LTX-2.5 are substantially more refined, reducing the fixes needed downstream and enabling generated shots to move to final with minimal additional post-production work.
With LTX-2.5, users can also expect greater stability across production workflows through:
- Native multishot generation. The model generates across several shots as one coherent output, holding the same characteristics from one frame to the next, reducing the risk of hallucinations and maintaining consistency and quality across the entire sequence.
- Sharper prompt understanding. A custom Gemma 4 backbone reads complex, multi-subject prompts more accurately, so a detailed direction resolves the way it was written and aligns closer to the creative intent on the first generation.
- HDR ACES in → HDR ACES out with the range intact. The ACES workflow allows for the loading of high dynamic range 16-bit linear data via EXR into LTX-2.5 in ACES and can run with LTX pipelines, while maintaining the range of the raw footage throughout the pipeline, producing outputs that can pass through a professional finishing pipeline.
Asteria, is an artist-led production company redefining cinematic storytelling, whose hybrid workflows built on LTX are setting the bar for this new form of filmmaking.
“Features like the ACES workflow are what excite us about LTX and its models. The LTX team is building genuine production workflows that integrate naturally into the hybrid environments we're creating at Asteria. LTX listens to filmmakers and delivers tools that work in real production. We can’t wait to start using LTX 2.5.”
Paul Trillo, co-founder & Executive Creative Director, Asteria
A Pipeline You Own
Because LTX-2.5 is open weights, professionals have full control over the model and its outputs, enabling teams to:
- Keep sensitive IP in-house. Run the model on your own hardware, so your footage, derivatives, the IP you feed it, and outputs you generate never pass through someone else's cloud.
- Reduce costs and accelerate turnaround. LTX models run at a fraction of the cost and generation time of closed, cloud-only tools.
- Fine-tuned results. The ability to adapt the model to your project through fine-tunes, LoRA and IC-LoRA workflows.
- Build a scalable pipeline. With LTX’s partnership with ComfyUI, LTX-2.5 is available from day one inside the node-based environment many teams already build in, so you can stand up a repeatable workflow from previsualization to final shot work rather than trying to organize a string of disconnected experiments.
LTX-2.5 is available today as open weights on Hugging Face, natively in ComfyUI, and through the LTX API for teams that want managed, production-grade generation. Go deeper on LTX-2.5 and how our open world models in the full announcement post.
