UniTemp: Unlocking Video Generation in Any Temporal Order via Bidirectional Distillation

A new framework called UniTemp enables video generation models to produce frames in any temporal order, overcoming a limitation that has kept autoregressive video diffusion models locked into forward-only generation [1][2]. Autoregressive video diffusion models have gained traction for long video generation in streaming settings, but their utility has been constrained by a strict forward temporal progression [1][2]. Practical video creation often demands flexibility — extending a clip backward from a known future frame, or filling in missing footage between two existing segments — workflows that forward-only models cannot support [2]. UniTemp, described in a paper submitted to arXiv on June 17, 2026, addresses this gap through a bidirectional distillation framework that trains a single autoregressive student model to generate video in arbitrary temporal directions [1][2]. The approach conditions on any combination of past and future frames at inference time, which the authors state improves controllability for both bidirectional extension and inbetween generation [2]. A central technical hurdle involved the Causal 3D VAE architecture common in video diffusion models. Because these variational autoencoders encode latents conditioned solely on past context, generating frames in reverse order produces visible inter-block discontinuities [2]. The researchers introduce blockwise anchor latents — auxiliary latents that restore the missing past context at block boundaries during backward generation — to resolve this artifact [2]. The paper reports that UniTemp maintains competitive performance on both short and long video generation benchmarks when measured against forward-only methods [2]. Beyond standard generation, the framework enables diverse workflows including bidirectional video extension, inbetween generation, looping video generation, scene transitions, and visual story generation [2]. The work appears on arXiv, the open-access repository that hosts electronic preprints across mathematics, physics, computer science, and related fields [6]. As of November 2024, arXiv was receiving approximately 24,000 article submissions per month and had surpassed two million total articles by the end of 2021 [6]. The UniTemp paper is accessible through the platform's abstract page, which includes arXivLabs integrations — a framework launched in 2020 that allows community collaborators to develop and share experimental tools directly on article record pages [4][5]. These integrations, displayed as tabs below the abstract, include bibliographic explorers, code finders, and recommender systems, all operating under guidelines that require partners to uphold arXiv's values of openness, community, excellence, and user data privacy [4].