Generative View Stitching

TL;DR

Autoregressive video diffusion can stay locally consistent but cannot condition on the future, so when a predefined camera trajectory loops back through a region the model has already painted, the rollout collides with its own scene and collapses. GVS replaces AR rollout with a training-free parallel sampling algorithm that denoises the entire sequence at once, exploiting a Diffusion Forcing model’s per-frame independent noise levels to extend single-image diffusion-stitching to video. Two ingredients close the gap: Omni Guidance conditions each segment on both past and future denoised neighbours, and a loop-closing schedule pairs the first and last segments during sampling. On the Impossible Staircase and other long predefined trajectories, AR baselines (even with RAG) diverge while GVS produces stable 120-frame loops.

Key claims

Camera-guided AR video diffusion collapses when a predefined trajectory revisits already-generated regions: the current step cannot see the future, so it paints a scene that later steps then have to collide with [§1, Fig. 1].
GVS is the first stitching-based sampling method for camera-guided video generation — it samples the full sequence in parallel by dividing it into overlapping segments and intertwining their denoising trajectories, so every frame is conditioned on both past and future neighbours [§1, §3].
The method is training-free and compatible with any off-the-shelf video diffusion model trained under Diffusion Forcing — the per-frame independent noise levels that Diffusion Forcing already supports are exactly the affordance stitching needs [Abstract, §3].
Prior stitching methods do not transfer: StochSync (panorama / mesh-texture stitching) lacks temporal consistency for video, and CompDiffuser requires a specially trained sequence diffusion model whose video version would be prohibitive to train [§2].
Omni Guidance — conditioning each segment’s denoising step on both past and future denoised context windows — is the headline ingredient that distinguishes GVS from prior stitching methods and is reported as essential for temporal consistency [Abstract, §3].
A loop-closing mechanism that pairs the first and last segments during the parallel denoising schedule yields long-range coherence on closed-loop trajectories (Impossible Staircase, closed circuits) where AR rollout, even AR+RAG, fails to close the loop [§1, Fig. 1].
Evaluated on a pretrained DFoT (Diffusion Forcing Transformer) video model with an 8-frame context window, GVS produces stable, collision-free 120-frame navigation videos on staircase and circuit trajectories where AR baselines collapse [Fig. 1].

Method

GVS treats the long-video generation problem as parallel denoising of an entire predefined-trajectory sequence rather than chunk-by-chunk AR rollout. The sequence is partitioned into overlapping segments; segments are denoised together, and at each step every segment’s context window is filled from both its past and future neighbours rather than only its causal past. The pretrained backbone is Diffusion Forcing — a sequence diffusion framework that gives each frame its own noise level — which means the sampling procedure can mix fully-denoised, partially-denoised, and noisy frames inside one context window without retraining. Two scheduling choices sit on top: Omni Guidance gates how strongly future-side context contributes to the current denoising step (so a segment can gradually commit to the scene its neighbours are painting), and a loop-closing schedule deliberately pairs the first and last segments during the parallel sweep so closed trajectories close. The procedure does not modify the backbone weights.

Results

On a pretrained DFoT 8-frame-context video model, GVS generates 120-frame navigation videos that are stable, collision-free, faithful to the predefined trajectory, frame-to-frame consistent, and close loops on a variety of paths including Oscar Reutersvärd’s Impossible Staircase [Fig. 1].
AR sampling on the same backbone diverges (collisions with the generated scene), is not faithful to the trajectory, and fails to close loops even when augmented with retrieval-augmented generation [Fig. 1].
StochSync (image-stitching baseline adapted to video) lacks temporal consistency on the same trajectories [§4.1].

(The paper reports headline-quality demonstrations on closed-loop and impossible-trajectory cases; quantitative tables sit alongside qualitative project-page video evidence at andrewsonga.github.io/gvs.)

Why it’s interesting

GVS is the cleanest articulation so far of a sampling-time alternative to the entire training-time long-AR-video lineage filed in this wiki — Self-Forcing, Context Forcing, PackForcing, MALT, FlowAct-R1, and Causal Forcing all attack the forgetting/drifting dilemma by changing how the AR student is trained (long-context teacher, partitioned cache, recurrent memory, error-recycling, fixed-init distillation). GVS instead leaves the AR-trained model alone and changes what gets fed into it at sample time — turning the AR backbone into a building block for a non-autoregressive sequence sampler. This makes it directly composable with much of the Autoregressive Video Generation cluster (you could in principle stitch a Context-Forcing-trained DFoT model) and gives the Camera-Controlled Video Diffusion cluster a third paradigm beyond explicit-trajectory adapters and reference-video conditioning: predefined-trajectory parallel sampling. It is also the first filed paper that frames the long-horizon problem as a future-conditioning problem rather than a memory/drift problem — closes loops by construction rather than by hoping memory is faithful enough.