LTX 2.5 vs LTX 2.3: What Actually Improved

The umbrella made the difference easy to see. LTX 2.3 began with it already open. LTX 2.5 showed the woman lifting and opening it, which was the action the prompt actually requested. In the longer test, 2.5 also held onto the rainy station, bicycle, bench, woman, and arriving train instead of gradually replacing the scene.

Those are encouraging results, not a universal verdict. We tested two prompts, and each version ran through its own compatible workflow. The comparison tells us how the complete 2.3 and 2.5 routes behaved on these jobs. It does not isolate one architectural change or predict every prompt.

Both versions ran on an NVIDIA RTX 3090 with 24 GB of VRAM. Outputs were 768 by 384 at 24 frames per second. The short case lasted five seconds and the multi-shot case lasted twelve seconds. Both used seed 250812. We retained the two-stage structure: generate at half resolution, upscale in latent space, then complete a short refinement stage before decoding video and audio.

The 2.3 result used the established LTX Director workflow. The 2.5 result used the official 2.5 transformer, Gemma 4 text encoder with LTX projection, audio VAE, convolutional video VAE, spatial upscaler, and distilled sigma schedule. This is therefore a workflow-level comparison, not a test in which only the transformer weights changed.

A fixed seed makes each route repeatable, but it does not create identical starting conditions across two different model stacks. Each stack interprets the random numbers through different learned weights, conditioning, and sampling code. Think of the seed as a numbered starting ticket, not a shared hidden video.

Fixed single-shot action
A woman in a mustard-yellow raincoat stands beside a red vintage bicycle.
She lifts a folded transparent umbrella, opens it once, then begins walking
the bicycle toward the camera. Preserve bicycle geometry, colors, identity,
rain, and exposure. Quiet rain and city ambience; no speech.

These are single observed runs, not a cold-cache speed benchmark. Loaded components and runtime state can change timing. The useful conclusion is modest: 2.5 did not introduce a dramatic slowdown in these tests.

Watch the umbrella first. The 2.3 clip contains movement, but it skips the opening action. In the 2.5 clip, the canopy unfolds while the red bicycle and rainy street remain recognizable. That is prompt adherence, not simply more motion.

Then watch the bicycle. Wheels, frames, and spokes are difficult because small geometric errors accumulate from frame to frame. Neither result is perfectly rigid, but 2.5 keeps the object readable while the requested human action takes place.

The longer prompt specified three four-second phases. First the woman walks the red bicycle toward a blue bench in a rainy outdoor station. Next she stops the bicycle and opens the umbrella. Finally a train arrives while the camera moves to a closer front view. Identity, object count, rain, colors, and exposure were locked throughout the text.

LTX 2.5 did not create three perfectly edited hard cuts. It evolved the angle and framing more fluidly. Even so, the requested chronology is visible and the scene does not abandon its own world.

• Environment: 2.5 keeps the wet outdoor platform; 2.3 drifts into a dry green corridor.
• Object continuity: the bicycle, bench, umbrella, and train remain part of one scene in 2.5.
• Editing: 2.5 treats the phases as a flowing camera sequence, not three clean editorial cuts.

A compatibility probe loaded the 2.5 transformer into a 2.3-era workflow. It failed before sampling because the older runtime expected a different text-conditioning structure. The model file itself was valid and GPU memory was available. The pieces simply did not speak the same interface.

A current ComfyUI runtime recognized the official Gemma 4 adapter and completed the two-stage audiovisual graph. This is why a model upgrade can resemble a software dependency upgrade: the transformer, text encoder, latent formats, VAE, scheduler, and node implementations form one contract.

Working LTX 2.5 contract
transformer: official 22B distilled INT8 ConvRot
text: Gemma 4 12B with LTX 2.5 projection
video VAE: LTX 2.5 convolutional VAE
audio VAE: LTX 2.5 audio VAE
upscaler: LTX 2.5 spatial x2
runtime: current ComfyUI with Gemma 4 LTX adapter support

The 2.5 multi-shot output contained audible rain and station ambience. The five-second single-shot outputs were technically valid stereo tracks but almost inaudible. The convolutional and DiffVAE video lanes produced the same very quiet audio, so the video decoder was not the cause.

The plotted mean level is a signal measurement, not a full perceptual-loudness score. Normalization can raise useful quiet ambience, but it cannot turn silence or meaningless noise into a good soundtrack. A practical pipeline should inspect duration, peaks, loudness, and silence, then either normalize, retry, or use a separate audio path. A successful render alone does not prove that both media streams are useful.

The quality decoder is viable at this output size, but the practical default should remain the convolutional VAE. It gives the worker more memory margin and already produced the best prompt-following examples in this experiment. DiffVAE belongs behind a quality option and needs broader testing at higher output sizes.

• Treat 2.5 as a complete stack. Verify the runtime, nodes, text adapter, VAEs, upscaler, and schedule together.
• Choose by intent. Keep low-level transformer and sigma choices out of the normal creative workflow.
• Check sound separately. Video quality and audio quality can fail independently even when they are generated together.
• Keep a known route available. Existing projects should not silently inherit a new model stack.
• Test more than attractive prompts. Dialogue, first-last-frame control, image guidance, and longer timelines can expose different failures.

Our two examples justify continued use and broader testing of LTX 2.5. They do not justify calling 2.3 obsolete. The more defensible decision is to make 2.5 available where the full stack and output checks are in place, then expand its role as a larger prompt suite confirms the result.

Official model, workflow, runtime, and license sources used to construct and interpret the benchmark.

• Lightricks LTX 2.5 model repository contains the official model card, files, and license for the tested release.
• Lightricks LTX-2 repository documents the official model family, inference code, and supported generation modes.
• Official ComfyUI-LTXVideo nodes provide the LTX-specific latent, conditioning, upscaling, and audiovisual workflow components.
• ComfyUI source repository contains the current Gemma 4 LTX adapter support required by the full-stack graph.