I don't think ComfyUI fully supports LTX 2.5 yet.
Maybe I'm missing something, or the official implementation is in the works, but one of the big hullabaloos of LTX 2.5 is DFR (Diffusion Fidelity Rendering). It's basically the first thing mentioned in the comfyui blog post about it. But...and this is awkward because I love comfy and everybody working on it.. it turns out, the full official DFR pipeline from Lightricks’ own LTX-2 GitHub repository is NOT actually implemented in ComfyUI.
But, wait! what if it's hidden inside the current implementation?...is what i asked chatgpt, when i went down this, couple of days deep rabbit hole. long story short, no it's not.
And therefore, I present to you actual native LTX-2.5 Spatial DFR in ComfyUI!
Here's a quick teaser from what I have working so far.
- T2V — fox / forest stress test
​
A low-angle wildlife documentary action-chase shot follows a red fox sprinting at full speed through a dense wet pine forest at dawn,backlit from the rising sun behind him, photographed with fully photorealistic natural detail. The camera races beside and slightly ahead of the fox at matching speed, keeping its head and upper body consistently framed while nearby ferns, wet grass and tree trunks sweep past with strong foreground and middle-ground parallax. The fox runs desperately, ears pulled back and body stretched through each stride; its paws kick wet leaves and droplets from the ground while its fur, whiskers and facial detail remain visible during the motion. Behind it, something enormous advances through the far forest but never enters the frame: first a distant tree top suddenly shudders, then a heavy branch snaps and splinters, and a moment later another heavy branch crashing through and falling with a violent impact noticeably closer, sending leaves and broken twigs outward as the unseen pursuer continues gaining ground. The fox briefly glances backward without slowing, then accelerates as the disturbance approaches, while layered morning mist and distant trees remain stable enough to preserve a strong sense of depth. Cold dawn light filters naturally through the canopy and catches moisture on fur, bark and vegetation. realistic anatomical deformations, Rapid paws stretching forward then crashing down on wet earth, the fox's breathing, distant cracking timber and deep approaching impacts dominate the soundscape. No cut, no visible monster, no fantasy styling or exaggerated debris explosion.
- I2V — cybernetic girl fidelity test against official reference I2V example
​
Use the provided start image as the first frame. The cybernetic figure slowly turns his head to the right, his glowing blue eyes scanning the horizon, mechanical joints in his neck whirring faintly. The camera follows his gaze, panning across the rooftop to reveal the city beyond: a river of light winding between dark towers, a flying vehicle gliding past between the buildings, its lights streaking. He watches it pass, then his eyes narrow slightly. The camera settles on his profile against the city glow, distant hover traffic humming, wind gusting across the rooftop. No text, no black frames.
So, the obvious question, how much slower is it? Yes, but it only became obvious to me when I started writing this post. Sorry, but when I was trying to get this to work I really didnt care how long it took. I was more focused on vram, memory loads, what was happening with sampling etc. But generally, I don't remember feeling a big difference at all. Anyway, I'm currently running tests with comfy-benchmark so I'll get back to you on this. :)
Oh I kept emphasizing "Spatial" DFR cause theres a Temporal upscaling in the full DFR pipeline. Already working on it, but I felt maybe this was a good enough milestone to share with you. It'll help make your LTX gens so much better.
So I'm finishing the final cleanup in repo and benchmark pass. Full update later today with all results, side-by-side comparisons, generated DFR keyframes, execution-time + VRAM benchmarks, workflows, and public access to the project.