July 2026Technical Report

Waypoint-1.5

A Real-Time Video World Model for Consumer Hardware

Rajit Rajpal^*Shahbuland Matiana^*Liew Wei Pyn^*Anmol Agarwal^*

Ryan CraigAndrew LappMithun HunsurSami BuGhanemScottie Fox

Aaron SandersCarson PooleIrene ParkDave RossiSpencer FrazierLouis Castricato

Overworld

^*Equal contribution

We present Waypoint-1.5, a real-time diffusion world model for interactive video generation on consumer-grade hardware. Unlike general video diffusion models, interactive world models (iWMs) must respond to dense user controls under strict latency and throughput constraints. Waypoint-1.5 is pre-trained on 100,000 hours of diverse, control-aligned video game data across hundreds of games, and generates playable video conditioned on full keyboard and mouse input. The model includes two resolution variants that run across a wide spectrum of consumer hardware. To characterize this unique setting, we distinguish rendered FPS, latent FPS, and control rate. We describe the data pipeline, architecture, training methodology, and runtime system behind Waypoint-1.5. We evaluate interactivity through latency, throughput, and inverse-dynamics based control recovery. Finally, we discuss the safety and ethics considerations unique to iWMs.

111

Latent FPS on RTX 5090

1.28B

Parameter single-stream causal DiT

>100K

Hours of training data

Full

Controller Conditioning with Mouse + Keyboard

Autoregressive Rollouts

Real-time gameplay generated by Waypoint-1.5 on a single consumer GPU, conditioned on keyboard and mouse input.

Architecture

Waypoint-1.5 is built on a single-stream causal Diffusion Transformer trained with per-frame independent noise levels (Diffusion Forcing), then distilled to a 4-step model via custom Self-Forcing DMD.

Training pipeline diagram — **Training Pipeline.** Per-frame independent noise levels under Diffusion Forcing training, with controller telemetry synchronized to each latent.

DiT architecture diagram — **Waypoint-1.5 DiT.** Single-stream causal DiT with weight-tied AdaLN noise conditioning and per-block MLPFusion controller injection.

Performance

Latent FPS across consumer GPUs at 512×1024 (high-res) and 256×512 (low-res) with INT8 quantization and BF16 baseline.

GPU	512×1024 INT8	512×1024 BF16	256×512 INT8	256×512 BF16
RTX PRO 6000 Blackwell	132.40	87.76	371.08	303.72
RTX 5090	111.40	66.28	346.64	277.52
RTX 4090	72.88	66.60	275.84	207.52
RTX 5070 Ti	59.68	39.96	270.48	178.56
RTX 3090 Ti	43.64	29.00	198.16	104.36
RTX 3090	37.36	21.60	188.44	116.12
RTX 5060 Ti	31.92	17.72	162.08	85.68
RTX 4060 Ti	28.32	16.40	126.84	78.68
RTX 3070	23.24	OOM	120.08	OOM
RTX 3060	15.76	8.80	81.88	36.76

BibTeX

If you use Waypoint-1.5 in your research, please cite:

@misc{overworld2026waypoint15,
  title         = {Waypoint-1.5: A Real-Time Video World Model for Consumer Hardware},
  author        = {Rajit Rajpal and Shahbuland Matiana and Liew Wei Pyn and
                   Anmol Agarwal and Ryan Craig and Andrew Lapp and
                   Mithun Hunsur and Sami BuGhanem and Scottie Fox and
                   Aaron Sanders and Carson Poole and Irene Park and Dave Rossi and
                   Spencer Frazier and Louis Castricato},
  year          = {2026},
  eprint        = {2606.XXXXX},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2606.XXXXX}
}