Waypoint 1.5#

Waypoint-1.5-1B is Overworld’s dense, autoregressive interactive video world model. FlashDreams integrates the published BF16 checkpoint as an image-established, keyboard/mouse-controlled V2 application with deterministic per-action metrics and live WebRTC presentation.

Upstream Waypoint 1.5 teaser from the Overworld model card; this is not a FlashDreams benchmark artifact.

Support summary#

Surface	FlashDreams support
Application slug	action2v-waypoint-1-5-1b through flashdreams-run-v2
Input modalities	RGB/RGBA first-frame image; keyboard and mouse buttons; relative mouse motion; ternary scroll-wheel direction
Output modality	Four RGB frames per action, native 1024x512 TCHW in the [-1, 1] range
Control modes	Live browser keyboard and mouse events
Output modes	WebRTC, MP4, null output, and optional per-action metrics
Precision and device	BF16 on CUDA; no FlashDreams quantized or CPU inference path
Not implemented	Text prompting, the 360P checkpoint, quantization, and multi-GPU execution

The checkpoint’s pinned configuration sets prompt_conditioning to null. Although the generic upstream APIs accept prompts for other models, this checkpoint-compatible FlashDreams path has no text encoder or prompt-conditioned weights and deliberately exposes no prompt argument. In the upstream runtime, this configuration leaves the prompt encoder uninitialized and calling set_prompt raises. The pinned safetensors artifact contains 393 tensor keys, with no keys for prompt or cross-attention modules. Consequently, this model does not accept a text prompt that can influence its output.

Requirements#

A CUDA-capable NVIDIA GPU with BF16 and PyTorch FlexAttention support.
The FlashDreams path was validated on one NVIDIA RTX PRO 6000 Blackwell Workstation Edition. The measured PyTorch peak allocation was 6.061 GiB, but this is not a minimum-VRAM guarantee and excludes non-PyTorch process memory.
The first run downloads the 3.72 GB BF16 Waypoint safetensors file and the separate Overworld-Models/taehv1_5 checkpoint into the Hugging Face cache.
ffmpeg on PATH is required for MP4 output.

Installation#

From the FlashDreams repository root:

uv sync --package flashdreams-waypoint --inexact

The package includes the model implementation and its Action2V adapter.

Running the model#

Run the application interactively in a browser:

uv run --no-sync flashdreams-run-v2 action2v-waypoint-1-5-1b \
    --mode webrtc --host 127.0.0.1 --port 8766 \
    -- --seed 464

Open http://127.0.0.1:8766/. Waypoint downloads its pinned default first frame when --first-frame is omitted. Arguments before the separator configure the V2 runtime; arguments after it configure Waypoint. To inspect all model arguments:

uv run --no-sync flashdreams-run-v2 action2v-waypoint-1-5-1b -- --help

Model and integration architecture#

The pinned checkpoint configuration and FlashDreams implementation agree on these model-facing invariants:

Dense 24-block autoregressive diffusion transformer, width 2048.
32 query heads, 16 K/V heads, and a four-times-width feed-forward network.
One model action is a 32-channel, 32x64 latent frame patchified into 512 spatial tokens.
Four rectified-flow Euler evaluations at sigmas 1.0, 0.9, 0.75, and 0.3; the terminal 0.0 evaluation commits clean cache state.
Control conditioning uses 256 button IDs, two mouse-delta values, and one ternary wheel value. Control fusion is present every third block.
Most blocks retain 16 latent actions densely. Blocks 3, 7, 11, 15, 19, and 23 use a 128-action horizon with every eighth historical action pinned.
TAEHV encodes four seed RGB frames into one latent action and decodes every generated latent action into four RGB frames.

The 128-action global horizon corresponds to 512 presented RGB frames, matching the context length stated by the upstream model card. FlashDreams presents the codec’s native 1024x512 canvas. The official world_engine client can instead resize 1280x720 input to that native canvas and resize decoded output back to 1280x720; FlashDreams intentionally omits those extra spatial resamples.

The upstream model card advertises a 1.2B parameter count. Independently, the pinned model.safetensors header contains 1,860,823,096 BF16 tensor elements across 393 tensors (3,721,694,304 bytes). These are different published-model versus serialized-checkpoint accounting figures; FlashDreams reports both rather than relabeling the upstream model.

The package-level design review contains component, class, use-case, and sequence diagrams:

Waypoint architecture design

Measured FlashDreams performance#

The final FlashDreams path was measured on 2026-08-26 using an RTX PRO 6000 Blackwell Workstation Edition (96 GiB), driver 595.84, PyTorch 2.12.1+cu130, CUDA 13.0, BF16 weights, the pinned validation first frame and control timeline, seed 464, native 1024x512 output, four denoise evaluations, and synchronous profiling. Actions 1-19 were warmup; actions 20-40 were the steady-state sample.

Measurement	Result
Mean action latency	68.791 ms per four generated frames
Median action latency	68.887 ms
p90 action latency	69.737 ms
Throughput derived from mean latency	58.15 generated RGB frames/s
Peak PyTorch CUDA allocation	6.061 GiB
Encoded output	164 frames: 4 seed + 40 actions x 4 frames

The declared 60 FPS is presentation timing, not a throughput guarantee. Overworld separately reports 56 FPS for its unquantized runtime on an RTX 5090; that result was not reproduced here and is not directly comparable across different GPU, runtime, and presentation stacks.

Parity and rollout validation#

FlashDreams loaded the same BF16 checkpoint and matched the pinned official world_engine implementation through one complete controlled action. The final comparison includes the four-step Euler solve and both cache commits:

Comparison	Mean absolute error	Max absolute error	Cosine similarity
Seed cache flow	0.010793	0.218750	0.999586
Clean latent after four steps	0.005682	0.033691	0.999869
Final cache flow	0.010511	0.265625	0.999487

The residual is consistent with BF16 execution through different compiled kernel boundaries. Review optimizations retained a byte-identical 40-action MP4. Additional inference completed 15 distinct scenes at 40 actions each and one 118-action rollout: 718 generated actions and 2,936 decoded frames in total, with complete finite metrics and exact frame accounting. This demonstrates execution, cache longevity, and scene coverage; it is not a qualitative gameplay or physical-accuracy score.

Complete validation record

Intended use and limitations#

Waypoint is suitable for research and prototyping around interactive video worlds, creative exploration, control-conditioned generation, and low-latency world-model systems. It is a generative model, not a physically grounded simulator.

Important limitations:

Long rollouts can drift, collapse, or become inconsistent.
Geometry, motion, object identity, and persistence can be unstable.
Outputs may reflect biases or unsafe patterns learned from training data.
The FlashDreams integration has no content-safety filter of its own.
It is not appropriate for safety-critical decisions, surveillance, high-stakes automation, or deployments that remove reasonable safeguards.
The current server accepts one WebRTC browser client per process. Multiple sessions may share model weights, but model execution is serialized.
Live results depend on browser event timing and are not reproducible from the application command line alone.

Review the upstream model card and world-model safety discussion before deployment:

Upstream model card

Upstream safety discussion

Provenance#

Model: Overworld/Waypoint-1.5-1B, revision 391f92827075edcf4a8b3c8a2ddae010698f8636, Apache-2.0.
Model SHA-256: b872ad07968bae082a120a29072e61a13565086f042384ad7fdb79a7b0c50994.
TAEHV: Overworld-Models/taehv1_5, revision a0253886b13b9c4c3bd224bd479be03f5988a3df.
Official parity implementation: Overworldai/world_engine at b3f1e725dedac17ccbfaf9ee37f5e068bb44bed4.
FlashDreams integration and documentation: Apache-2.0.