# Autoregressive Video Generation

PixVid converts a pretrained latent video diffusion model into a frame-wise pixel-space generator: it keeps the diffusion transformer, replaces the VAE decoder with a convolutional detailer head, and finetunes from Wan2.1-14B. It reaches VBench totals of 85.09 at 960p and 84.78 at 480p, above Wan2.1's 84.05 and PixelUMM's 83.72, and keeps 98% of source motion on fast clips versus 83% for an identically finetuned latent model.

PixVid converts a pretrained latent video diffusion model into a frame-wise pixel-space generator: it keeps the diffusion transformer, replaces the VAE decoder with a convolutional detailer head, and finetunes from Wan2.1-14B. It reaches VBench totals of 85.09 at 960p and 84.78 at 480p, above Wan2.1's 84.05 and PixelUMM's 83.72, and keeps 98% of source motion on fast clips versus 83% for an identically finetuned latent model.

Interactive world models have been limited by memory that grows with stream length: temporal context runs out of GPU memory, and explicit 3D memories slow down as redundant primitives accumulate frame by frame. L3G-World replaces that with a compact latent 3D Gaussian memory that spawns Gaussians only in newly observed regions, refines existing ones with residual updates, and stores a VAE latent feature in each Gaussian so the memory renders directly into the diffusion transformer's latent space without re-encoding. On 481-frame loop-closure trajectories from DL3DV and RealEstate10K, full L3G-World reaches PSNR of 16.00 and 18.36, and with few-step distillation it generates at 9.02 FPS on a single H100.

Interactive world models have been limited by memory that grows with stream length: temporal context runs out of GPU memory, and explicit 3D memories slow down as redundant primitives accumulate frame by frame. L3G-World replaces that with a compact latent 3D Gaussian memory that spawns Gaussians only in newly observed regions, refines existing ones with residual updates, and stores a VAE latent feature in each Gaussian so the memory renders directly into the diffusion transformer's latent space without re-encoding. On 481-frame loop-closure trajectories from DL3DV and RealEstate10K, full L3G-World reaches PSNR of 16.00 and 18.36, and with few-step distillation it generates at 9.02 FPS on a single H100.

In autoregressive video diffusion, writing generated frames into key-value context and denoising the current frame previously shared one set of parameters; SGF+ gives each role its own parameters while keeping them coupled through causal attention, and still trains both with the original generation objective. On VBench at 5s, 60s, and 240s horizons, SGF+ scores above Self Forcing and Self Gradient Forcing on most quality and long-horizon consistency metrics in both framewise and chunkwise generation; trained only on 5s clips, it generates continuously for up to 24 hours without long-video fine-tuning.

In autoregressive video diffusion, writing generated frames into key-value context and denoising the current frame previously shared one set of parameters; SGF+ gives each role its own parameters while keeping them coupled through causal attention, and still trains both with the original generation objective. On VBench at 5s, 60s, and 240s horizons, SGF+ scores above Self Forcing and Self Gradient Forcing on most quality and long-horizon consistency metrics in both framewise and chunkwise generation; trained only on 5s clips, it generates continuously for up to 24 hours without long-video fine-tuning.

SemanTok is a flexible-length video tokenizer that feeds frozen DINOv2 features into its encoder and adds lightweight heads that reconstruct those features from each retained token prefix alone, letting a 201M autoregressive model match or beat a VideoFlexTok model 3.4 times its size and improving semantic alignment at every noise level, including pure noise.

SemanTok is a flexible-length video tokenizer that feeds frozen DINOv2 features into its encoder and adds lightweight heads that reconstruct those features from each retained token prefix alone, letting a 201M autoregressive model match or beat a VideoFlexTok model 3.4 times its size and improving semantic alignment at every noise level, including pure noise.

FlashForward publishes the in-flight KV cache already computed by each ordinary denoising forward for reuse by later chunks and complements it with sparse clean anchor KV generated ahead of time for long-range structural guidance, removing cache-update-only forwards; with up to four GPUs it runs 1.16–1.69x faster than HiAR and 1.42–2.92x faster than Self-Forcing for 16 FPS videos of 20 seconds or longer across 1.3B and 14B backbones at 480p and 720p, and the 1.3B model at 480p scores 0.838 on VBench-1.0 while remaining stable at 20s, 35s, and 65s.

FlashForward publishes the in-flight KV cache already computed by each ordinary denoising forward for reuse by later chunks and complements it with sparse clean anchor KV generated ahead of time for long-range structural guidance, removing cache-update-only forwards; with up to four GPUs it runs 1.16–1.69x faster than HiAR and 1.42–2.92x faster than Self-Forcing for 16 FPS videos of 20 seconds or longer across 1.3B and 14B backbones at 480p and 720p, and the 1.3B model at 480p scores 0.838 on VBench-1.0 while remaining stable at 20s, 35s, and 65s.

EditWorld is a video world model that accepts streaming editing instructions and reference images during autoregressive generation through Gated Causal Attention, a Sparse Context mechanism, joint autoregressive and bidirectional training, and a dedicated data synthesis and annotation pipeline, and it introduces WBench-Editing with roughly 150 cases of 240–480 frames each, achieving the best overall score of 73.8 and an editing score of 80.0 on that benchmark.

EditWorld is a video world model that accepts streaming editing instructions and reference images during autoregressive generation through Gated Causal Attention, a Sparse Context mechanism, joint autoregressive and bidirectional training, and a dedicated data synthesis and annotation pipeline, and it introduces WBench-Editing with roughly 150 cases of 240–480 frames each, achieving the best overall score of 73.8 and an editing score of 80.0 on that benchmark.