WanToFight: Real-Time Generative Game Engine for Multi-Player Combat Interaction

Bang Zhang; Guangyuan Wang; Li Hu; Peng Zhang

REVIEW 3 major objections 3 minor 1 cited by

WanToFight: Real-Time Generative Game Engine for Multi-Player Combat Interaction

T0 review · 3 major / 3 minor · reviewed 2026-07-15 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read WanToFight generates real-time two-player King of Fighters '97 combat from keyboard input, uniting multi-player control, physical interaction, and adversarial play in one generative engine.

desk verdict Abstract-only systems claim of first real-time dual-control generative KOF engine; interesting if true, but multi-player binding is asserted without evidence. read the letter →

arxiv 2607.12592 v1 pith:XD43ZIWC submitted 2026-07-14 cs.CV

Li Hu , Guangyuan Wang , Peng Zhang , Bang Zhang This is my paper · ORCID

classification cs.CV

keywords generativegameenginemulti-playercombatvideodiffusiontransformerreal-timeinferenceKingofFighterskeyboardcontrolplayerassociationmodeldistillation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper presents WanToFight, a generative game engine that simulates complete two-player matches of The King of Fighters '97 in real time from ordinary keyboard input. Earlier generative game systems covered only single-player first-person scenes or non-real-time cooperative play; none jointly delivered multi-player control, real-time speed, complex physical interaction, and adversarial combat. WanToFight builds on a 1.3-billion-parameter video diffusion transformer, adding a streaming generator with rolling memory, a module that visually links each keyboard stream to the correct fighter, and a gated keyboard injection path trained by gradually moving from single-player to full adversarial gameplay. After distillation and decoder pruning, the system sustains 30 frames per second at 512×384 resolution on one high-end GPU for an entire match. If the approach holds, modern video generators can act as interactive multi-player game engines rather than only as passive video synthesizers.

What carries the argument

A streaming autoregressive video generator based on the Wan-1.3B diffusion transformer with block-causal attention and a rolling KV cache; a visually grounded Player Association module that binds each player's keyboard signal to a character identity; and a gated, locally causal keyboard injection module trained under a single-player-to-full-gameplay curriculum, then distilled with a pruned VAE decoder for real-time speed.

What would settle it

Run a full two-player match and check whether each fighter consistently responds only to its assigned keyboard stream across punches, combos, jumps, and knockdowns; any sustained identity swap, frozen character, or crosstalk where one player's keys move the wrong fighter falsifies the multi-player claim.

Watch

Extended reading notes

Core claim

WanToFight is the first generative game engine to jointly deliver multi-player control, real-time inference, complex physical interaction, and adversarial gameplay, by simulating two-player KOF '97 from keyboard input at 30 FPS on a single GPU for a complete match.

Load-bearing premise

That the visual Player Association module and the gated keyboard injection reliably keep each player's controls attached to the correct fighter throughout a match, without identity swaps or control crosstalk that would break adversarial combat.

Editorial extensions

If this is right

Full two-player fighting-game matches can be driven end-to-end by a generative video model without a conventional game engine.
Separate keyboard streams can be bound to distinct on-screen characters and produce coherent adversarial physics in real time.
Distilled diffusion models can sustain 30 FPS interactive video generation at modest resolution on a single high-end GPU for multi-minute sessions.
The same association-plus-injection recipe could support other multi-player interactive titles once trained for them.
Generative engines become candidates for interactive entertainment and training simulators that previously required hand-authored physics and animation systems.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The Player Association idea may generalize to any multi-agent video task where multiple control streams must map to distinct visual entities without identity swaps.
Success at adversarial combat would suggest generative models can internalize reaction and counter-play dynamics, not merely replay cooperative motion.
If the single-player-to-full-gameplay curriculum is the decisive training step, similar staged curricula could unlock other complex multi-agent settings such as team sports or racing.
Long-match consistency of the rolling KV cache will determine whether the approach scales beyond short fighting rounds.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

This is an abstract-only systems paper claiming the first generative game engine that jointly does multi-player keyboard control, real-time inference, complex physical interaction, and adversarial combat (KOF '97) at 30 FPS 512×384 on one RTX 5090 for a full match. That joint capability is the punchline. If it holds, it is a real within-subfield advance over prior single-player or non-real-time generative engines.

What looks new is the integration, not the individual pieces. They start from Wan-1.3B, add streaming block-causal attention with rolling KV cache, a four-step DMD-distilled student plus pruned VAE for the speed claim, a visually grounded Player Association module, and gated locally causal keyboard injection trained under a single-player-to-full-gameplay curriculum. The engineering story is coherent on paper: distillation and pruning for latency, association plus gated injection for dual control, curriculum to stage the hard multi-agent interaction. Credit where due—the abstract is clear about the gap it targets and names concrete modules and a concrete operating point.

The soft spot is exactly the one the reader and stress-test flag, and it is load-bearing: we have no evidence that Player Association plus gated injection actually binds each keyboard stream to the right character and sustains coherent adversarial physics under streaming generation. No architecture details, no identity-swap or crosstalk rates, no ablations of association or curriculum, no dual-control metrics over full matches, no failure modes. The 30 FPS full-match claim is also just stated. Free parameters (DMD steps, cache policy, curriculum schedule) and the invented modules are uninspectable from the abstract alone. Circularity is low—this is an empirical systems claim—but soundness is currently low because the central multi-player result is asserted rather than shown.

Who it is for: people building interactive generative video / game engines who care about dual control and real-time adversarial settings. A serious referee should see the full paper (methods, numbers, ablations, videos). I would not desk-reject on the abstract; the claimed joint capability is important enough to evaluate properly. I would not cite or bring to reading group until there is evidence the binding works. Send it to review if the full text materializes with real measurements; otherwise it stays a claim.

Referee Report

3 major / 3 minor

Summary. The manuscript introduces WanToFight, a generative game engine that aims to simulate real-time two-player The King of Fighters '97 gameplay from dual keyboard streams. Built on a Wan-1.3B video diffusion transformer, the system comprises (i) a streaming autoregressive generator with block-causal attention and a rolling KV cache, (ii) a visually grounded Player Association module that binds each keyboard stream to a character identity, and (iii) a gated, locally causal keyboard injection path trained under a single-player-to-full-gameplay curriculum. A four-step DMD-distilled student with a pruned VAE decoder is claimed to sustain 30 FPS at 512×384 on one RTX 5090 for a complete match. The central claim is that this is the first generative game engine to jointly deliver multi-player control, real-time inference, complex physical interaction, and adversarial gameplay.

Significance. If the system works as stated, the contribution is substantial for generative interactive media: prior generative game engines have largely been limited to single-player first-person or non-real-time cooperative settings, and jointly solving dual-control binding, streaming real-time generation, and adversarial physical combat would close a clear capability gap. The engineering stack (block-causal streaming, association module, gated injection, DMD distillation, pruned VAE) is a concrete systems package that, if validated, would be useful to the community. The abstract-only record, however, does not yet let a referee confirm that those capabilities hold under dual control over full matches.

major comments (3)

The load-bearing multi-player claim rests on the Player Association module plus gated keyboard injection (trained under the single-player-to-full-gameplay curriculum) reliably binding each concurrent keyboard stream to the correct on-screen fighter and sustaining coherent adversarial physical interaction under streaming block-causal generation. The abstract asserts this binding but reports no identity-swap / control-crosstalk rates, dual-control accuracy over full matches, ablations of association or curriculum, or failure-mode characterization. Without those measurements the multi-player and adversarial claims cannot be assessed.
The real-time performance claim—30 FPS at 512×384 on a single RTX 5090 for a complete match via a four-step DMD-distilled student and pruned VAE—is stated without measurement protocol (wall-clock vs. model-only, cache warm-up, resolution/cache policy, match length distribution) and without baselines or latency breakdowns. For a systems paper whose novelty includes real-time dual-control generation, this evidence is required to support the central claim.
The priority claim that WanToFight is the first system to jointly combine multi-player control, real-time inference, complex physical interaction, and adversarial gameplay is not supported in the abstract by a comparative table or explicit positioning against named prior generative engines. A referee cannot verify novelty or scope of the gap without that comparison and the corresponding dual-control metrics.

minor comments (3)

The abstract names free design choices (DMD step count, spatial/temporal operating resolution and cache policy, curriculum schedule) without even order-of-magnitude values; a short quantitative teaser would help readers gauge the operating point.
Notation for the three modules (Player Association, gated locally causal keyboard injection, streaming block-causal generator) is introduced only by name; a one-line functional description of inputs/outputs for each would improve abstract clarity.
No dataset scale, character/move coverage, or training compute is mentioned; even a brief clause would situate the empirical scope.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: systems paper claims empirical real-time multi-player generation capability, not a derivation that recovers its inputs by construction.

full rationale

WanToFight is an engineering/systems abstract describing a generative game engine (streaming Wan-1.3B-based distilled model, Player Association, gated keyboard injection, curriculum, DMD distillation, pruned VAE) that claims to run two-player KOF '97 from keyboard input at 30 FPS. There is no mathematical derivation chain, no fitted constants renamed as predictions, no uniqueness theorem imported from the authors, and no self-definitional loop (X defined via Y then used to derive Y). The only available text is the abstract; it asserts components and a performance claim without equations that reduce outputs to inputs by construction. Self-evaluation without external dual-control benchmarks is ordinary for systems abstracts and is not circularity under the stated criteria. Score 0 with empty steps is the correct honest finding.

Assumptions & free parameters 3 free parameters · 4 assumptions · 2 invented entities

Abstract-only systems paper. Free parameters are architectural and training choices that must be set for the claimed runtime and control fidelity. Axioms are standard ML/systems assumptions plus domain assumptions that video diffusion can stand in for a fighting-game physics/state engine. Invented entities are the named modules that carry the multi-player binding and control story; none have independent evidence outside the claimed system.

free parameters (3)

DMD distillation step count = 4 steps
Abstract specifies a four-step DMD-distilled student; step count and distillation hyperparameters are design choices that determine the real-time claim.
Spatial/temporal operating resolution and cache policy = 512x384, 30 FPS target
512×384 output, rolling KV cache, and block-causal attention windowing are chosen operating points that trade quality for 30 FPS; not derived from first principles.
Single-player-to-full-gameplay curriculum schedule
Curriculum stages and mixing ratios for keyboard injection training are free training design choices asserted to enable dual-player control.

assumptions (4)

domain assumption A pretrained Wan-1.3B video diffusion transformer is a suitable backbone for interactive fighting-game frame generation.
The system is built on Wan-1.3B; suitability for dual-control adversarial combat is assumed rather than derived.
domain assumption Block-causal attention with a rolling KV cache preserves temporal consistency under continuous streaming for a full match.
Streaming real-time generation depends on this architectural assumption stated in the abstract.
standard math Standard diffusion training, DMD distillation, and VAE decoding mathematics apply without modification that would invalidate the latency claim.
Background ML methods treated as given.
ad hoc to paper Visually grounded association can uniquely and stably map two concurrent keyboard streams to two on-screen fighters throughout combat.
Core multi-player premise introduced for this system; not a standard theorem.

invented entities (2)

Player Association module
purpose: Bind each player's keyboard signal to a character identity in the generated video.
Named as a new component required for multi-player control; abstract gives no external falsifiable handle beyond the system's own demos.
Gated locally causal keyboard injection module
purpose: Inject dual keyboard controls into the generative stream while preserving local causality and enabling curriculum training.
Introduced to solve control injection for adversarial two-player play; evidence is internal to the claimed system.

how reviews work

0 comments

Cite this review

Pith. "Pith review of WanToFight: Real-Time Generative Game Engine for Multi-Player Combat Interaction." pith.science (2026). https://pith.science/paper/XD43ZIWC

@misc{pith2026260712592,
  author       = {Pith},
  title        = {Pith review of: WanToFight: Real-Time Generative Game Engine for Multi-Player Combat Interaction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XD43ZIWC}},
  note         = {Machine review of arXiv:2607.12592}
}

read the original abstract

We present WanToFight, a generative game engine that simulates real-time, two-player The King of Fighters '97 (KOF~'97) gameplay from keyboard input. Prior generative game engines target either single-player first-person settings or non-real-time cooperative scenarios; multi-player control, real-time inference, complex physical interaction, and adversarial gameplay have not been jointly addressed. WanToFight closes this gap with three components built on the Wan-1.3B video diffusion transformer: a streaming autoregressive generator with block-causal attention and a rolling KV cache; a visually grounded Player Association module that binds each player's keyboard signal to a character identity; and a gated, locally causal keyboard injection module trained with a single-player-to-full-gameplay curriculum. A four-step DMD-distilled student paired with a pruned VAE decoder sustains 30FPS at 512x384 on a single NVIDIA RTX 5090 over the duration of a complete match. To our knowledge, WanToFight is the first generative game engine to combine multi-player control, real-time inference, complex physical interaction, and adversarial gameplay in one system.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

MASS: Multiplayer World Models with Authoritative Shared State
cs.CV 2026-08 conditional novelty 6.0 of 10

MASS predicts an explicit typed world state once per tick, then renders any number of camera views from that same state, improving state recovery and cross-view consistency on a Snake benchmark.

Pith tools

Reviewed July 15, 2026 · model on record in the stance chip above.