Edit with references.
Keep what matters.

Regenerate a source with identity, composition, and geometry preservation—optionally borrowing details from a second reference.

Source · Picture 1The image to edit

Optional · Picture 2Style, garment, object…

Describe the final editPicture 1 is always the source. If supplied, call the optional reference Picture 2.

CanvasRecipe

Source fidelity 75%Prompt preservation strength—not a diffusion denoise slider.

Measured GPU-time billing · $0.30 current estimate12-step default · native ~0.98 MP cap

REAL H3 EXAMPLE

Real REF2VA editClassifier-verified example output

Classifier-routedEvery source and output is classified; adult inputs use the compatible H3 lane and adult outputs stay out of public search.

Shared model residencyText-to-image shares FL2VA, Qwen, and VAE weights with video. REF2VA is loaded only when editing needs it.

Measured, then drainedFinal price follows actual execution; the RunPod endpoint scales back to zero after its queue empties.

Edit with references.Keep what matters.