Image to Image¶
Repaint an input image under a new prompt and export a PNG. Ideal for stylization, touch-up, or partial rewrites.
Minimal example¶
Key parameters¶
| Parameter | Type | Default | Notes |
|---|---|---|---|
image |
str |
required | path or URI to input image |
prompt |
str |
required | rewrite prompt |
strength |
float ∈ (0, 1] |
0.8 |
rewrite intensity; lower keeps more of the original |
negative_prompt |
str? |
None |
negative prompt |
preset |
fast/balanced/quality/low-vram |
balanced |
see Presets |
num_inference_steps |
int? |
preset | explicit step override |
guidance_scale |
float? |
preset | CFG |
seed |
int? |
random | reproducibility |
Supported models¶
- Mainstream:
sd15,sd21,sdxl-base-1.0 - Refiner:
sdxl-refiner-1.0(typically used as a second pass oversdxl-base-1.0)
Run omnirt models --task image2image for the full list.
Common recipes¶
- Light stylization:
strength=0.4,preset=balanced - Heavy rewrite:
strength=0.85,preset=quality - SDXL two-pass polish: draft with
sdxl-base-1.0, then runimage2imageonsdxl-refiner-1.0atstrength=0.3
Troubleshooting¶
Warning
- Invalid input dimensions — validate first:
omnirt validate --task image2image --model <id> --image <path> --prompt "…" strengthoutside (0, 1] — caught at validation time- Color shift / ghosting — usually
strengthtoo high or too few steps; trystrength=0.55, preset=balanced