Failure modes in sub-second causal events

Sub-second events compress cause and effect into a very small number of frames — typically 3-8 frames at 24fps. The model has little temporal space to repair inconsistencies once the event begins. An error in frame 2 of a 5-frame sequence cannot be corrected by frame 5 without violating temporal coherence.

Glass, water, sparks, and smoke are particularly difficult because they involve many small particles that must still obey a coherent global event. Each particle has its own trajectory, but all particles must collectively respect conservation laws and material properties. The model must generate a distribution of particle behaviors that appears random at the individual level while remaining deterministic at the aggregate level.

The fundamental challenge is that these events require the model to simulate physics forward in time with high precision. Unlike slow-moving objects where small errors can be masked by subsequent motion, fast events leave no room for error accumulation. The first frame of the event sequence must be nearly perfect for the subsequent frames to remain plausible.

Additionally, fast events often involve phase transitions (solid to liquid, liquid to gas, intact to fragmented) that require the model to understand material properties at a level deeper than surface appearance. A glass object must not only look like glass before shattering but must behave like glass during the shatter event.

Pre-impact fragments: Small pieces of the breaking object appear in the frame before the impact event occurs. This suggests the model is anticipating the outcome and leaking future state into the present frame.

Splash shape reversal: Water or particle trajectories that reverse direction mid-motion, violating conservation of momentum. This typically occurs when the model attempts to 'correct' an earlier error by adjusting subsequent frames.

Ignition without fuel continuity: Fire or sparks appear without a visible source of ignition or combustible material. The model generates the visual effect of combustion without establishing the necessary preconditions.

Debris without source: Fragments or particles appear in the scene without any visible origin object. This indicates the model is generating secondary effects without maintaining the causal chain from primary event.

Velocity discontinuities: Objects or particles that change speed or direction instantaneously between frames, creating a strobing or teleporting effect. This violates the expectation of smooth motion under constant forces.

Material property violations: Objects that behave inconsistently with their apparent material — glass that bends before breaking, water that maintains shape like a solid, smoke that falls instead of rising.

These errors are easy to miss in single-frame review but become obvious when the clip is scrubbed frame by frame. We recommend a 'causal chain validation' protocol: for each frame in the event sequence, verify that the visual state could have arisen from the immediately preceding frame through physically plausible transformations.

Automated detection is possible for some artifact types. Velocity discontinuities can be flagged by computing optical flow between consecutive frames and identifying motion vectors that exceed expected maximum velocities. Pre-impact fragments can be detected by comparing object presence across frames and flagging elements that appear before their causal trigger.

However, many failure modes require human judgment. Determining whether a splash shape is physically plausible requires understanding fluid dynamics; identifying material property violations requires knowledge of the object's composition and behavior under stress.

Use prompts with a clear trigger (the cause), a measurable reaction (the effect), and a stable camera position. Avoid unnecessary cinematic motion when the goal is causal evaluation — camera movement introduces additional variables that can mask or compound event-level errors.

A good benchmark isolates the event before adding complex lighting, camera moves, or character interaction. Start with simple scenarios: a glass on a table, a drop of water falling, a match being struck. Once the model demonstrates competence on isolated events, gradually increase complexity by adding interacting elements.

We recommend reporting not just aggregate success rates but per-event-type performance. A model might excel at rigid body dynamics (glass shatter) while failing at fluid dynamics (water splash). Granular reporting enables targeted improvement and helps users understand the model's specific capabilities and limitations.

Finally, include negative examples in the benchmark: events that should NOT occur. A model that generates plausible physics for requested events but also generates spurious events (ignition without cause, fragments without impact) should receive lower scores than a model that generates fewer but more accurate events.