Guide: Getting smooth panning shots without the warping effect.

Initial testing of Luma Dream Machine revealed a persistent issue: generated panning shots often exhibit severe spatial warping, where background elements stretch unnaturally as the "camera" moves. This is a classic failure mode in video diffusion models, but it can be mitigated with careful prompt engineering and parameter tuning.

Through systematic iteration, I've identified a protocol that significantly increases the probability of a smooth, coherent pan. The core principle is to **explicitly anchor the scene** and **decompose the motion**.

### Key Prompt Structure

Avoid single-action prompts like `"panning shot over a mountain range."` Instead, use a structured scene description followed by a precise, separated camera instruction.

```
A static, wide-shot landscape of a misty mountain range at sunrise. The mountains are detailed and stable in the frame. [camera slowly pans left to right]
```

### Critical Parameters (API/Advanced Settings)

* **Motion Bucket ID:** This is crucial. Lower values (e.g., `64-128`) produce slower, more stable motion. High values (`>200`) for fast pans almost guarantee warping.
* **Generation Length:** Always generate at least `5` seconds (`frames=164`). Shorter clips often lack the stabilization phase.
* **Negative Prompting:** Explicitly add terms like `"warping," "liquid motion," "melting background," "distorted perspective," "blurry motion."`

### Recommended Workflow

1. **Establish a Baseline:** Generate your target scene *without* camera motion first. Verify the model can render the subject stably.
2. **Introduce Simple Motion:** Apply a very slow, single-axis pan (`camera slowly pans right`) with `motion_bucket_id=100`.
3. **Iterate and Refine:** If warping occurs, reduce the `motion_bucket_id` and simplify the background. Complex, repetitive textures (forests, bricks) warp more easily.
4. **Post-Processing:** Even successful pans may need stabilization. A light pass through a tool like `ffmpeg` with `vidstab` can correct minor jitter.

The model appears to prioritize maintaining object coherence over geometric precision during motion. Therefore, prompts describing simple, distant backgrounds with a clear focal point (e.g., `"a lone tree on a hill, panning shot"`) consistently outperform complex, detailed interiors where every element can warp independently. It's a trade-off between scene complexity and motion fidelity.