NVIDIA Chrono Edit

Frequently Asked Questions

What is NVIDIA Chrono Edit and when should I use it?

Chrono Edit is a physics-aware image editor that performs edits by reasoning through a short imagined motion between the original and the requested change, preserving geometry, lighting, reflections, and shadows. Use it for realistic object swaps, material changes, additions/removals, and physically consistent edits.

How does "imagined motion" keep pose, lighting and shadows consistent?

The model simulates a brief temporal transition from the source to the edited state, which helps it respect contacts, camera angle, pose, and scene illumination, leading to more coherent outputs.

What inputs are required and which formats are supported?

A single image and a concrete instruction prompt are required. Use a clean, well-lit photo in JPEG or PNG for best fidelity.

How should I write prompts for physics-aware edits?

State the physical change and the invariants to preserve. Example: ‘replace the plastic mug with a clear glass half-filled with tea; keep the same angle and morning light.’ Avoid generic style fluff; be explicit about conservation (pose, camera angle, lighting).

Can I preserve the original camera angle and subject pose?

Yes—explicitly instruct to keep camera angle and pose. Chrono Edit prioritizes consistency unless you ask for lighting or viewpoint changes.

Can I change lighting or material without breaking realism?

Yes. State lighting/material changes clearly (e.g., ‘switch to afternoon sunlight’ or ‘change plastic to brushed metal’). If not specified, the model preserves original lighting.

How do I ensure reproducibility and iterate quickly?

Use a fixed integer seed. The same image + prompt + seed yields the same result. Iterate by fixing seed and making small prompt adjustments.

Which output formats are supported?

JPEG and PNG. Choose JPEG for compact distribution and PNG for lossless editing.

Should I crop to a region before editing faces or text?

For faces or text areas, cropping closer to the region can improve fidelity and sharpness in the edited result.

Why did my edit drift from the request and how can I fix it?

Tighten the prompt to specify what to change and what to preserve; fix the seed to compare iterations; ensure input quality is high (sharp, well-lit, minimally compressed).

Does it return results in BASE64 or via sync response?

Both are supported at the API level (BASE64 and sync return). In this app, results are returned as URLs by default.

Are there input quality limitations or best practices?

Input quality sets the ceiling. Avoid soft, noisy, or heavily compressed sources; provide clean lighting and clear subjects for best realism.