Lore

Prompt chaining on one image for consistent frames

Chaining several prompts against the same image so that the subject's face and pose stay consistent while everything else changes: environment, framing, outfit, era and lighting.

Use

The host presents this as a way to build a set of related frames, for example the frames of a vertical video, without regenerating the subject from scratch each time.

Building blocks

Each step is a small instruction against the current image, and the model carries the identity forward. When identity or detail slips, correct it as described in generate-then-correct-hybrid-workflow.

Compare Multi-reference mode with character sheets and an environment, which holds identity with character sheets instead of a chained image.

Chained zoom example: face to eye macro

Zoom prompts can be chained, with each step taking the previous result as input:

  1. Zoom to the face.
  2. Zoom to a macro of the left eye.
  3. Zoom closer on the eye.

The host reports that the last step shows every tiny vein in the white of the eye and a small water droplet on the eyelashes. The detail is generated at each step. It is not enlarged from the first image.

Because each step works from the previous frame, the frames stay consistent with each other. The sequence can serve as a set of stills for a video zoom-in; see prompt-based-lens-zoom for why this differs from a digital crop.