Kling 3.0 and Luma: A Reproducible Physics-Stress Test
On this page
Quick answer
Neither Kling nor Luma should be called “melting-free” without a controlled, repeatable evaluation. AI video models generate plausible frames; they do not guarantee a physically correct simulation. Results can change with the model version, mode, duration, aspect ratio, seed, reference image, and prompt.
Flowith did not run a controlled Kling-versus-Luma benchmark for this page. Instead, this guide defines a test that a production team can run on the current versions available in its accounts.
What this test measures
Separate the visible failure types rather than compressing them into one subjective quality score:
- shape stability: does an object’s silhouette or volume drift?
- contact: do hands, tools, floors, or containers intersect incorrectly?
- trajectory: does direction or speed change without a visible cause?
- material behavior: do water, cloth, smoke, or fragments move plausibly?
- temporal continuity: do textures and identities persist across frames?
- camera interaction: does subject motion remain coherent while the camera moves?
The result is a failure profile, not a universal winner.
Test setup
Use the currently available models on the Kling and Luma Dream Machine first-party products. Before generating, record:
- test date and region;
- product surface and plan;
- exact model and mode shown in the interface;
- resolution, aspect ratio, duration, and frame rate;
- prompt, negative prompt, reference media, and seed if exposed;
- number of attempts and any discarded generations;
- credits consumed and queue time.
Run the same prompt and reference inputs in both tools. Use at least three attempts per scene; one lucky or unlucky output should not decide the comparison.
Six stress scenes
1. Rigid-body impact
Prompt a ceramic mug falling from a table onto a hard floor. Check acceleration, contact timing, bounce, breakage, and fragment persistence. Flag unexplained shape changes separately from implausible motion.
2. Liquid and container
Pour water from a clear pitcher into a glass. Watch the stream, surface response, transparency, container edges, overflow, and the relationship between liquid volume and fill level.
3. Cloth in airflow
Use a scarf attached at one point and moving in a steady wind. Check edge continuity, folds, material identity, self-intersection, and whether the attachment point stays fixed.
4. Fire and smoke
Generate a stationary campfire with a locked camera. Review source continuity, smoke direction, turbulence, dissipation, lighting on nearby objects, and background stability.
5. Hand-object contact
Show a hand lifting a filled glass. This combines anatomy, contact, a rigid object, and liquid. Count finger-glass intersections, grip discontinuities, container deformation, and water behavior.
6. Moving camera
Repeat the falling-mug scene with a tracking camera. This isolates whether camera motion increases subject jitter, background warping, orientation changes, or trajectory errors.
Review worksheet
Have two reviewers inspect anonymized outputs without knowing the model. For each attempt, record failure counts and timestamped notes in the six categories above. Then reveal the model and compare:
| Question | Evidence |
|---|---|
| Which failure occurred? | Timestamp and frame |
| Was it visible at normal playback? | Yes / no |
| Would it require regeneration or editing? | Action and estimated time |
| Did all attempts fail similarly? | Attempt count |
| Did a setting change affect the result? | Exact setting |
Keep raw outputs and prompts with the worksheet. Do not average unlike defects into a precise “physics score” unless the weighting was defined before review.
How to make a workflow decision
Choose based on the scenes that resemble the actual production brief. A model that handles liquid better in this test may still be a poor fit if the project depends on character continuity, audio, edit controls, licensing, or predictable cost.
Recheck current product documentation before a purchase. Model names, modes, limits, and prices change, and a result from one release should not be generalized to another.