HY-World 1.5 (WorldPlay): A Systematic Framework for Interactive World Modeling with Real-Time Latency and Geometric Consistency

TL;DR

HY-World 1.5 (codenamed WorldPlay) is Tencent Hunyuan’s open-source streaming video-diffusion world model: a single image or text prompt seeds an interactive rollout that responds to keyboard/mouse input at 24 FPS with long-term geometric consistency. Four design pillars carry the system: a Dual Action Representation for robust action control, a Reconstituted Context Memory that dynamically rebuilds context from past frames with temporal reframing to keep geometrically important long-past frames accessible, WorldCompass — a novel RL post-training framework targeting action-following and visual quality on the autoregressive backbone, and Context Forcing, a memory-aware distillation method that aligns memory context between teacher and student to enable real-time speeds without error drift. The release ships HunyuanVideo-1.5-based (8B) and WAN-based (5B) variants, including bidirectional, autoregressive, RL-post-trained, and distilled checkpoints, plus full training and inference code.

Key claims

Discrete keyboard/mouse inputs are mapped to a Dual Action Representation that supports both first- and third-person perspectives across real-world and stylized environments [README §Introduction; §Inference — Camera Trajectory Control supports w/s/a/d movement and up/down/left/right rotations with action-duration syntax].
Reconstituted Context Memory dynamically rebuilds context from past chunks at every 16-frame next-chunk-prediction step, using temporal reframing to keep geometrically-important but long-past frames in scope — explicitly aimed at the memory-attenuation failure mode of fixed-context AR video [README §Introduction; §Inference Pipeline].
WorldCompass is an RL post-training framework designed to directly improve action-following and visual quality of long-horizon autoregressive video models — released as a separate code module with its own paper [arXiv:2602.09022] [README §News; worldcompass/README.md].
Context Forcing distillation aligns memory context between teacher and student to preserve the student’s capacity to use long-range information while reaching real-time speeds without error drift [README §Introduction].
Real-time inference at 24 FPS, 480p, 125-frame generations is achievable on a single GPU with sequence parallelism (sp=8 → 28 GB; sp=4 → 34 GB; sp=1 → 72 GB) [README §System Requirements].
The full HY-World 1.5 system outperforms prior baselines on both reconstruction and human evaluation: at the long-term horizon, PSNR 18.94 / SSIM 0.585 / LPIPS 0.371 / Rdist 0.332 / Tdist 0.797 vs Gen3C’s 15.37/0.431/0.483/0.357/0.979 and Matrix-Game-2.0’s 9.57/0.205/0.631/2.125/2.742 [README §Evaluation table].
HY-World 1.5 is the only model in its comparison table that is both real-time and the long-term-consistency leader — Gen3C/ViewCrafter/CameraCtrl/SEVA/VMem/GameCraft are non-real-time; Matrix-Game-2.0 is real-time but trails on all five long-term metrics [README §Evaluation table].
The Context Forcing distillation contributes measurably to long-term consistency: ablating it drops long-term PSNR 18.94 → 16.27, SSIM 0.585 → 0.425, LPIPS 0.371 → 0.495, Tdist 0.797 → 0.991 [README §Evaluation, “Ours (w/o Context Forcing)” row].
The release includes four open-source autoregressive checkpoints — bidirectional, AR, AR+RL, AR+distilled (4 steps) — plus a fifth (AR+RL+distilled) to be released, all on the HunyuanVideo-1.5 8B base, plus a separate WAN-based 5B lightweight pipeline for small-VRAM GPUs at compromised quality [README §Model Checkpoints; §Inference].
Training data is mined from 100+ AAA games (HY-World 1.0 lineage), the same data philosophy as Hunyuan-GameCraft [README §News, January 6, 2026 entry — “training code for WorldPlay-8B model”; Slack note: “mostly trained with AAA games (similar to Gamecraft2)”].
The technical report and accompanying arXiv research paper (2512.14614) were released December 17, 2025; a follow-up WorldCompass RL paper (2602.09022) was released March 8, 2026 [README §News].

Method

The pipeline is a chunkwise autoregressive next-chunk prediction model: given a single image or text prompt, the model emits 16-frame video chunks conditioned on per-step user actions. For each chunk, Reconstituted Context Memory dynamically reconstitutes the conditioning context from past chunks rather than maintaining a fixed FIFO buffer — past frames are temporally reframed so geometrically-relevant moments stay accessible even if temporally distant.

Backbone. HunyuanVideo-1.5 8B (recommended, with stronger action control and long-term memory) or WAN 5B (lightweight, compromised). The HunyuanVideo path additionally uses Qwen2.5-VL-7B-Instruct as the text encoder, ByT5-small + Glyph-SDXL-v2 for glyph encoding, and FLUX.1-Redux-dev’s SigLIP as the vision encoder.

Action control. Camera trajectory is specified either as an intuitive pose-string DSL (e.g. 'w-3, right-1, d-4' for “forward 3 latents, yaw right 1, strafe right 4”) or via custom JSON files generated by generate_custom_trajectory.py. Internally these are converted to the Dual Action Representation that drives the diffusion conditioning.

RL post-training (WorldCompass). A separate post-training stage on the AR checkpoint that directly optimizes action-following and visual quality, yielding the ar_rl_model variant. Code lives in worldcompass/.

Distillation (Context Forcing). A memory-aware distillation procedure that aligns the memory-context distribution between teacher and student so the student preserves long-range information utilization. Produces the 4-step ar_distilled_action_model for real-time inference; an ar_rl_distill_model (RL + distilled) is the only checkpoint still in the TBD list.

Engineering. Optional FP8 GEMM via AngelSlim + DeepGEMM, SageAttention (required for the WAN pipeline), and Flash-Attention for accelerated HunyuanVideo inference. Sequence parallelism (sp=1/4/8) trades VRAM for throughput.

Results

Quantitative evaluation reported in the README compares HY-World 1.5 against seven prior systems on real-time / short-term / long-term metrics:

Model	Real-time	Long-term PSNR ↑	Long-term SSIM ↑	Long-term LPIPS ↓	Long-term Rdist ↓	Long-term Tdist ↓
CameraCtrl	❌	10.09	0.241	0.549	0.733	1.117
SEVA	❌	10.51	0.301	0.517	0.721	1.893
ViewCrafter	❌	9.32	0.271	0.661	1.573	3.051
Gen3C	❌	15.37	0.431	0.483	0.357	0.979
VMem	❌	12.77	0.335	0.542	0.748	1.547
Matrix-Game-2.0	✅	9.57	0.205	0.631	2.125	2.742
GameCraft	❌	10.09	0.287	0.614	2.497	3.291
Ours (w/o Context Forcing)	❌	16.27	0.425	0.495	0.611	0.991
Ours (full)	✅	18.94	0.585	0.371	0.332	0.797

[README §Evaluation]. The headline result is the combination of real-time (24 FPS) and long-term leadership across all five long-term reconstruction metrics — every prior real-time entrant (Matrix-Game-2.0) is dominated on all five, and every prior long-term leader (Gen3C) is non-real-time. The Context Forcing ablation shows that distillation contributes the bulk of the long-term-consistency win (PSNR 16.27 → 18.94, Tdist 0.991 → 0.797) rather than being a pure efficiency trick. A human-evaluation figure is referenced but not reproduced in numerical form in the README.

Why it’s interesting

This is the open-source counterpart to the closed real-time-interactive WFM flagships (Project Genie: Experimenting with infinite, interactive worlds, Introducing Runway Gen-4.5) and the immediate successor to Tencent’s own Hunyuan-GameCraft: High-dynamic Interactive Game Video Generation with Hybrid History Condition — same data philosophy (AAA-game-mined frames), same goal (interactive game-video generation), but extending from PCM-distilled 6.6-FPS chunk generation to a full streaming 24-FPS rollout with explicit long-horizon memory and an RL post-training stage. The “Context Forcing” name collides with Context Forcing: Consistent Autoregressive Video Generation with Long Context in the Context Length / Quality Trade-off in Video Generation cluster but means something different here — there it’s a functional KV-cache partitioning at inference; here it’s a memory-aware distillation procedure that preserves long-range context utilization in the student. The WorldCompass RL stage is the wiki’s first datapoint of Reasoning RL applied to an autoregressive video world model, sitting alongside Identity-GRPO: Optimizing Multi-Human Identity-preserving Video Generation via Reinforcement Learning (identity-preserving video RL) as a second domain-specific RL post-training recipe for video DiTs. Filed today because the AAA-games training recipe + open weights make this the canonical open-source competitor to closed real-time WFMs and the most directly comparable system to a Luma in-house interactive world model.