Resolve the interaction.
The engine updates fighter state, resolves hits, and applies the active rules.
Fight World Model
Customizable combat through
visual interaction proxies
Execute the rules. Render the fight.
1 University of Oxford2 Independent Researcher3 Imperial College London4 Tsinghua University5 Nanyang Technological University
Executable state / Conditional videoScroll to explore
00The idea
Every hit has a consequence.
A protected fighter can react without losing health. Lifesteal can heal the attacker. FWM connects these executable rules to the frames you see.
Put outcomes on a timeline.
A visual proxy aligns actions, reactions, and health with each fighter and frame.
Give the fight its appearance.
A shared video renderer follows the proxy and the character or arena references.
01Screening room
Many ways to stage a fight.
Characters, arenas, and skill effects in FightWorld Corpus.
The videos in this screening room and the corpus gallery are edited training samples. FWM inference comparisons appear in the paper figures below.
Appearance / Identity
A new face in the arena.
Appearance / Environment
Take the fight somewhere else.
Effects / Casting windows
Set the skill in motion.
02Under the image
One canvas. A shared timeline.
Attempted actions, transient reactions, and persistent health have separate marks. Fighter identity ties them together, even when the fighters switch sides.
Actions and skill windows.
Horizontal stripes mark punches. Vertical stripes mark kicks. Coverage encodes strength. A magenta block follows the caster during a skill window.
Reactions and health.
A brief color inversion marks a reaction. Bar length independently records health. Edited rules can change one without changing the other.
Inside the model
The engine executes the rules, the compiler rasterizes the records, and a video diffusion transformer renders the continuation. Shared identity embeddings bind references to the intended fighter.
03Executable rules
Change the rule. Recompute the fight.
The engine reruns the same initial state and submitted actions under edited rules. The resulting proxy changes. The trained renderer stays the same.
Source: the paper's rule intervention study. Health MAE is in percentage points of maximum health. Joint rule fidelity requires the applicable reaction and health checks to pass together. These are reported evaluation results.
04FightWorld Corpus
A broader vocabulary of combat.
Native gameplay and edited videos share the same training format, with interaction proxies, text descriptions, and available appearance references.
112,618Training samples
52Skill effects
187Character appearances
15Arenas
The sample archive.
74 selected corpus clips. Open a sample to compare its video, proxy, and reference.
05Evaluation
The whole interaction has to agree.
Joint fidelity checks actions, entity assignment, timing, reactions, and persistent health together. All complete-information encodings below receive the same executed records.
98.48%
Joint interaction fidelity
99.79%
Correct hit reactions
0.15 pp
Health mean absolute error
| Control representation | Move ↑ | Attack ↑ | Skill ↑ | Reaction ↑ | Health MAE ↓ | Joint ↑ |
|---|---|---|---|---|---|---|
| Spatial proxy + state tokens | 94.48 | 91.23 | 82.42 | 87.17 | 1.23 | 63.28 |
| Spatial proxy + text | 96.72 | 93.63 | 92.43 | 92.24 | 2.78 | 76.35 |
| FWM interaction proxy | 99.78 | 99.24 | 99.06 | 99.79 | 0.15 | 98.48 |
Appearance across side switches
Identity binding raises character adherence on side-switch clips from 53.68% to 98.27% under matched reference inputs and training budgets. The identity follows the fighter as its screen position changes. Source: the paper's identity binding study.
Evaluated in fixed-camera, two-fighter 2D combat with a single video backbone. Transfer to other games, moving cameras, 3D worlds, and more fighters remains to be evaluated.