NVIDIA ChronoEdit — Physics-aware Image Editing
ChronoEdit performs edits by reasoning through a short “imagined motion” between the original and the requested change. That temporal reasoning helps keep geometry, lighting, shadows, and contacts consistent in the final image.
Highlights
- Consistency-first edits: pose, lighting, reflections, and shadows stay believable.
- Works with one input image plus a natural-language instruction.
- Reproducible results with seed control; quick iteration for creative teams.
- Production-friendly outputs in JPEG, PNG, or WEBP; optional BASE64 and sync return.
Parameters
-
prompt:
Describe the change you want, not general style fluff. Example: replace the plastic mug with a clear glass half-filled with tea; keep same angle and morning light.
-
image:
Source image URL or upload. Use a clean, well-lit photo for best fidelity.
-
seed:
Integer for determinism. Same prompt + image + seed gives the same result. Use −1 to randomize.
-
output_format:
jpeg for compact distribution, png for lossless editing, webp for web delivery.
How to Use
- Upload the source image.
- Write a concrete prompt that states the target edit and what must be preserved.
- Set seed if you need reproducibility.
- Choose output_format.
- Run and review. If something drifts, tighten the prompt and fix the seed.
Prompting Tips
- Be explicit about conservation: keep camera angle, keep morning sunlight, keep the person’s pose.
- State the edit as a physical change: swap, add, remove, change material, change color.
- If lighting must change, say it clearly; otherwise the model will preserve the original light.
- For faces or text areas, crop tighter to the region before editing for crisper results.
Pricing
Per run: $0.02
Quality Notes
- Input quality sets the ceiling. Soft, noisy, or heavily compressed sources limit realism.
- For large pose or object changes, iterate with fixed seed and small prompt adjustments.