Features improvement
Feature Request: Multi-reference character consistency (like Seedance's 50-image reference)
LTX-2.5's IC-LoRA currently focuses on motion/structure transfer (Canny, Depth, Pose) rather than direct multi-reference character consistency. Competing models like Seedance 2.5 support up to 50 reference images for strong character consistency across shots.
It would be great if LTX's next version included native support for multiple reference images (face/character sheets) to lock character identity across generations, similar to what Seedance offers. This would make LTX much stronger for AI filmmaking and character-driven storytelling workflows.
Additionally, Seedance 2.5 generates a single native clip up to 30 seconds long (with multi-round extensions up to 180 seconds in beta) at 4K resolution in one pass. It would be great to see LTX push its native single-generation length closer to this range as well, since longer coherent single-pass clips reduce the need for manual stitching in production workflows.
Thanks for the great work on LTX-2.5!
Feature Request: Multi-reference character consistency (like Seedance's 50-image reference)
LTX-2.5's IC-LoRA currently focuses on motion/structure transfer (Canny, Depth, Pose) rather than direct multi-reference character consistency. Competing models like Seedance 2.5 support up to 50 reference images for strong character consistency across shots.
It would be great if LTX's next version included native support for multiple reference images (face/character sheets) to lock character identity across generations, similar to what Seedance offers. This would make LTX much stronger for AI filmmaking and character-driven storytelling workflows.
Additionally, Seedance 2.5 generates a single native clip up to 30 seconds long (with multi-round extensions up to 180 seconds in beta) at 4K resolution in one pass. It would be great to see LTX push its native single-generation length closer to this range as well, since longer coherent single-pass clips reduce the need for manual stitching in production workflows.
Thanks for the great work on LTX-2.5!
This is actually great feature to have i hope they implement it soon . Really love ltx 2.5 .
I agree! Native support for multiple face and character references would be a valuable addition, especially for film restoration.
Being able to supply clear references from different angles could help preserve an actor’s identity across shots while keeping the original expressions, movement, and performance intact.
I created LTX Reference Restorer, a locally hosted web app to help organize this workflow: source clips, actor references, shot guides, versioned render jobs, and side-by-side comparisons.
GitHub: LTX Reference Restorer
The project includes example workflows using public-domain films such as The Great Train Robbery, St. Louis Blues, and The Hitch-Hiker, with documented Library of Congress sources.
It’s an early preview: the reference-management and baseline-upscaling workflows are tested, while actual LTX inference remains unverified. My goal is to make reference-guided restoration easier to prepare, evaluate, and improve. Having stronger multi-reference support built directly into LTX would make this kind of workflow much more useful.
Thanks to the team for continuing to develop LTX!