Lore

Generate, then correct: alternating model output with Photoshop edits

The host treats Nano Banana Pro output as a strong starting point, not a finished result. His best results come from alternating generation with manual correction.

Flaws the host reports

Two ways to fix them

  1. Re-prompt with the right references. Add or swap reference images so the model restores the outfit, face or setting. See Garment reference image for outfit transfer.
  2. Use conventional Photoshop tools. The host names blur, curves and masks. A blur restores depth of field, curves correct color and grading, and masks combine layers (see Subject masking and background replacement).

Order of work

Generate, inspect at full size, correct the specific defect, then feed the corrected image back in for the next generation step. This works well with Prompt chaining on one image for consistent frames and depends on the 4K layers from Nano Banana Pro in Photoshop via the Luminel plug-in.

Model weaknesses and their Photoshop fixes (from the Nano Banana Pro for Pros video)

The host's workflow shows the same pattern throughout: generate with the model, then correct what it reliably gets wrong with ordinary Photoshop tools.

| Model weakness | Photoshop fix | |---|---| | Weak or missing depth of field on a sharp background | Field Blur with pins for depth of field | | Rectangular far-background patch left unblurred | Local Gaussian blur patch fix | | Face loses detail in a later pass | Mask restoration of lost face detail | | Color drifts from the background reference | Curves color correction for reference drift | | Outfit drifts from the sketch | Re-feeding the sketch to correct outfit drift (a re-prompt, not a Photoshop edit) |

The order of steps matters more than any single fix. The host openly treats correction rounds as part of the workflow.

Hand-finishing generated character art

Manual correction and painting over generated character art, used to fix the palette and unify the hybrid 3D/watercolor look.

This follows the merged design from Batch-of-four critique and trait merge and feeds the character inputs described in Asset-first approach.