Replace the decoder without touching the rest

Pruna AI's PrunaVAED decoder cuts the final rendering time in half on LTX-2.3. A reduction in video memory that makes clip production smoother.

A video generation model does not work directly on pixels. It manipulates a compressed representation, the latent, which a final brick called a decoder transforms back into visible images. This final step is memory-intensive, and it is precisely what PrunaVAED—a lightweight version of the LTX-2.3 decoder by Pruna AI—targets.

The appeal lies in its interchangeable nature. The encoder and the latent format remain identical, meaning this brick replaces the original one without modifying anything upstream. On an H100 GPU in bfloat16, the publisher measures a five-second 720p decoding time reduced from 841 to 500 milliseconds, and a ten-second 1080p decoding time cut from 4903 to 2356 milliseconds—representing speedups between 1.67x and 2.08x depending on length and resolution, with peak video memory usage reduced by about half.

The method combines pruning, retraining, and distillation. The decoder remains strictly identical to the original during its initial stages, with the reduction only starting afterward, through a 25% and then a 50% channel reduction in subsequent blocks. As a result, the decoder goes from 407 million to 345 million parameters. Visual discrepancies measured across two hundred clips remain low, with a SSIM above 0.98 and a dedicated metric for stability between consecutive frames.

The limitations are stated upfront. The measurements come from a single GPU, with a batch size of one, on only two LTX-2.3 pipelines, and the gains vary depending on hardware and settings. The output is not bit-exact compared to the original decoder, does not replace either the generator or the denoiser, and does not accept latents from other sources. Finally, as a derivative of LTX-2.3, distribution is governed by the LTX-2 community license and its usage restrictions, rather than an open-source license.