Prompt: Edit the input character image into one clean 3x3 cinematic storyboard grid of keyframes that an image-to-video model can interpolate.

Layout: exactly 3 columns and 3 rows of equal 16:9 panels, thin neutral borders, no outer padding. A tiny 01–09 marker sits in each panel's top-left corner tab, reading left to right, top to bottom; markers never cover the character, face, hands or props, and there is no other text.

The beat: the nine panels are one continuous ~15-second moment with the same subject(s) in the same location, not nine separate scenes. Follow the user's action if they give one. Otherwise infer one simple micro-sequence from the pose, props, gaze, outfit and setting, and match the source's energy: action sources may dodge, dash, draw or strike; calm sources turn, glance, reach, step, breathe or react, with a subtle peak rather than an explosive one. Use only props that are visible.

Progression: 01–03 set up and prepare, 04–06 change decisively and hit the peak (motion, contact, reveal or emotional turn), 07–09 carry the follow-through and aftermath. Panel 09 ends in a new pose, clearly different from panel 01, unless the user asks for a loop.

Character lock: every panel keeps the reference's face, hair, eyes, outfit, accessories, body type and art style. A single character stays the protagonist; several keep their roles, spacing and scale. No clones, mirror doubles, alternate outfits or new focal characters; imply tension through gaze, shadow, sparks, wind or light instead. Keep faces readable: no extreme back views, masks or heavy occlusion. Never stretch or squash the character to fill a wide panel; crop, change camera distance or extend the background, and the full body need not appear in every panel.

Continuity: location, time of day, weather, color grade, key light and landmarks stay consistent. Each panel continues directly from the one before: no time skips, resets or teleports. Carry over pose logic, prop state, eyeline and screen direction.

Cinematography: vary framing where it fits (wide, medium, close-up, extreme detail on hand, eye or prop, high or low angle, over-shoulder, profile, push-in) along one implied camera path, without abrupt scale jumps between neighbors. Anime keyframe language: clear silhouettes, weight transfer, hair and cloth motion, expressive faces, motion blur only where speed demands it. Clean, strongly lit, coherent perspective.

Avoid: identity or art-style drift; chibi or western-comic redesigns; speech bubbles, sound-effect text, logos or any text beyond the markers; warped anatomy or hands; distorted or unequal panels.

The final frame shows them leaning across the table to kiss each other Model: GPT Image 2.5 Sunburst v2.5