You prompt pouring water into a glass and the water flows sideways. Or a flag waving in wind and the cloth ripples in patterns no real fabric would make. Or a ball rolling down a hill and it pauses, accelerates wrong, or rolls uphill for a beat. The action is recognizable but physically impossible.
Fastest fix: don’t ask the model to render physics it can’t. Reframe the shot so the hard physics happens off-screen (cut around it), and use a model with stronger physics for what’s left. As of June 2026 the strongest physics models are Kling 3.0 (best on water, fire, smoke, collisions), Sora 2 (still excellent on simulation), and Veo 3.1 (best on cloth and hair, good on gravity). For high-stakes commercial shots, composite real physics footage over an AI background.
AI video models learn from clips, but they do not run a physics engine. They have seen thousands of examples of water pouring, yet they do not “know” water always flows down — they statistically reconstruct what pouring looks like. For simple, common cases that holds up. For edge cases it breaks dramatically.
Which bucket are you in?
| Symptom | Most likely cause | Go to |
|---|---|---|
| Water/smoke/fire moves wrong; cloth ripples unnaturally | Model has no real physics for fluids/cloth | Step 1, Step 5 |
| Action has many stages (tilt, pour, fill, stop) and one stage breaks | Action is too multi-stage for one generation | Step 1, Step 2 |
| You started from a still image and the motion looks invented | Reference frame doesn’t show the action in progress | Step 3 |
| Same shot works on Kling/Sora/Veo but fails on your tool | Weaker model on physics | Step 5 |
| Object is rare (lab gear, niche sports kit, specialty tools) | Thin training data for that object | Step 1, Step 6 |
| Physics looks fine at the start, drifts apart by the end | Clip too long; error accumulates | Step 4 |
Common causes
Ordered by hit rate, highest first.
1. The model has no physics engine
Every current video model (Kling, Sora 2, Veo, Runway, Pika, Seedance) has limited physical understanding. They reliably struggle with:
- Fluid dynamics (water, smoke, fire)
- Cloth dynamics (flags, capes, hair)
- Multi-object collisions (balls, dominoes)
- Gravity-dependent motion (falling, jumping, dropping)
- Friction-dependent motion (sliding, rolling on surfaces)
How to spot it: your shot involves one of these phenomena. Expect issues and plan around them.
2. The action is complex or multi-stage
pours water from a kettle into a cup until full involves tilt, flow, accumulation, level rise, and stop. Five separate physics moments all have to land in one generation.
How to spot it: the action has more than one physical stage.
3. The reference image doesn’t match the action
You uploaded a still of a kettle held over a cup and asked for “pouring.” The model has to interpolate the entire pour from a single frame — too much extrapolation, too much room for error.
How to spot it: your source frame doesn’t show the action already in progress.
4. The tool or model version is weak on physics
As of June 2026, Kling 3.0, Sora 2, and Veo 3.1 handle physics noticeably better than Runway Gen-4.5, Pika, and older versions of any model. Smaller or older models are worse.
How to spot it: the same shot works on Kling/Sora/Veo but fails on your current tool.
5. The object is one the model rarely sees
Unusual objects (specialized tools, science equipment, niche sports gear) have less training data, so the model invents physics for them.
How to spot it: the object in the shot is uncommon.
6. The clip is too long for stable physics
Like other drift problems, a longer clip means more accumulated physics error.
How to spot it: physics is fine at the start and falls apart by the end of a long clip.
Shortest path to fix
Step 1: Reframe the shot to avoid the physics
The single best fix is to not ask the model to do physics it can’t.
# Replace physics-heavy shots with surrogate shots
# Bad - direct physics
"water pouring from kettle into cup, water level rising"
# Good - implied physics via cuts
1. Hand tilting kettle (no water visible in the flow)
2. Cut to a cup with the water level high (already filled)
3. Add a "water pouring" sound effect for sensory connection
# Bad - direct collision
"ball bounces off wall and rolls"
# Good - surrogate
1. Ball flying toward the wall (just before contact)
2. Cut - sound effect of impact
3. Ball already rolling awayEditors call this “cutting on the action.” The audience fills in the physics you never rendered.
Step 2: Generate static-state shots, edit in the motion
For impossible-to-render physics, generate the start, middle, and end as near-still states and let editing imply the motion between them:
1. Generate the "before" state (full kettle held over an empty cup)
2. Generate the "during" state (water mid-pour, no realistic motion required)
3. Generate the "after" state (kettle put down, cup full)
4. Edit them together with sound designThis is a workaround pros use even on platforms with good physics.
Step 3: Use reference-video / motion control where the tool supports it
If you have real footage of similar physics, trace its motion instead of letting the model imagine physics:
# Runway Gen-4.5 control tools (as of June 2026)
- Motion Brush 3.0: paint up to 5 regions, each with its own
direction + speed vector (horizontal, vertical, proximity)
- Video-to-video: restyle real footage so the underlying motion
comes from the source clip, not from physics prediction
- Act-One: drive a character's performance from a reference video
# Kling and Veo also offer motion-reference / video-to-video modesThis bypasses the physics-prediction step: the model follows the path you supplied rather than guessing.
Step 4: Shorten the clip
# Cap clip length
- Physics error compounds; 2-3s is the sweet spot for hard shots
- 5s+ will show visible physics drift on difficult casesIf you need length, generate in short physics-safe segments and stitch them, rather than asking for one long clip.
Step 5: Switch to a stronger physics model
# Strongest on physics (as of June 2026)
- Kling 3.0 (Kuaishou) - best on water, fire, smoke, and collisions
- Sora 2 / Sora 2 Pro (OpenAI) - excellent simulation; the
standalone Sora app AND the in-ChatGPT Sora button were both
removed Apr 26 2026, so the only remaining access is the
Sora 2 API (which itself sunsets Sep 24 2026)
- Veo 3.1 (Google) - best on cloth and hair, strong on gravity
and collisions
# Improving
- Seedance 2.0 (ByteDance) - good, but fluids/multi-body still approximate
# Weaker on hard physics
- Runway Gen-4.5, Pika, and older versions of any model
# Note: even the best model fails on truly complex physicsMost production teams route between two or three models by scene type rather than relying on one.
Step 6: For commercial work, composite real physics into the AI footage
# Hybrid workflow
1. Generate the AI footage WITHOUT the problematic physics
2. Shoot or source real footage of the physics (water, fire, etc.)
3. Composite the real physics over the AI background in
After Effects / DaVinci Resolve Fusion / Nuke
4. Use rotoscoping or masks to integrate the layersThis is the gold-standard approach for ads, music videos, and any high-stakes physics scene.
How to confirm it’s fixed
- Play the clip at 0.25x speed. Slowing it down exposes physics breaks (water changing direction, objects accelerating wrong) that read as “fine” at full speed.
- Loop the segment 3-4 times. If your eye snags on the same frame each loop, that’s where the physics fails — re-cut or regenerate that beat.
- Check the start and the end of long clips separately. Drift usually appears in the second half.
- Show one person who didn’t make the shot. If they can’t name the artifact, the surrogate or composite did its job.
Prevention
- Keep a running list of physics shots your model handles and shots to avoid.
- For physics-heavy needs, plan the cut-based workaround up front, not in production.
- Test physics shots with a stand-in / proxy before committing to a full render.
- For high-stakes work, plan from day one to composite real physics over the AI background.
FAQ
Which AI video model has the best physics in 2026? For fluids, fire, smoke, and collisions, Kling 3.0 leads as of June 2026. Sora 2 is still excellent on overall simulation, and Veo 3.1 is best for cloth and hair. There is no single winner, so production teams pick by scene type.
Is Sora still available? The standalone Sora app and website shut down on April 26, 2026, and the Sora button was pulled from ChatGPT on the same date. Sora 2 now lives on only through the Sora 2 API, which is itself scheduled to be discontinued on September 24, 2026. If you depend on it, plan a migration to Kling or Veo.
Will prompting harder fix bad physics? Mostly no. Adding “realistic physics, accurate gravity” rarely overrides the failure because the model has no physics engine to invoke. Reframing the shot, switching models, or compositing real footage works far more reliably than prompt wording.
Why does it look fine at the start and then break? Physics error accumulates frame by frame. The longer the clip, the more drift compounds, so the second half of a 5s+ shot is where collapses usually show. Keep hard-physics clips to 2-3s.
Can motion control or reference video fix this? Yes, often. Runway’s Motion Brush 3.0 (up to 5 painted regions with direction and speed) and video-to-video modes, plus Kling’s and Veo’s motion-reference modes, make the model follow a supplied path instead of predicting physics, which removes the most common failure.