AI Champion 2026 — Re:Member

A Demo of a Platform That Brings Memories Back

Re:Member: Living Memory (‘Re:Member 살아있는 기억’) is a platform proposal: it draws a person’s memory scenes out through conversation with a digital human, brings those scenes back as images and video, and keeps them as a life record. We entered it in AI Champion 2026 as a four-person team of two professors and two researchers, and I handled the demo video. The name changed three times in ten days: from Re:Memory to Link: Memories of Generations (「링크: 세대의 기억」), then to Re:Member the day before submission.

The video shows exactly what the platform does. When an elderly speaker says it was near the sea, a scene appears. When they correct it to a rocky shore, the scene changes. When they say it was a place like a harbor, a harbor with moored boats appears. The cycle — speak a memory, the AI offers a scene, fix what’s off, make it again — is edited in the same order as the lines.

Re:Member system diagram — user device, digital human engine, AI conversation engine, story generation engine, video production pipeline, storage and distribution — The six-stage system diagram in the proposal. The diagram itself was drawn with GPT Image 2.

Nine Days of Production

Date	Work
Apr 13	Proposal draft
Apr 16	Scene image generation: six 3×3 sheets and one single image (56 min)
Apr 17	Organized video prompt schemas (Veo, Kling, Seedance, Nano Banana)
Apr 18, early morning	Generated 22 clips and a first cut in Premiere Pro, 2:21 (01:46–04:03)
Apr 22–23	Added 5 dialogue shots and 4 Kling clips, made the system diagram
Apr 23–24	Five After Effects versions, upscaling and overlays, final 2:58 render, submission

Stage	Tool	Input → Output
Images	GPT Image 2	Scene prompt → 5504×3072 3×3 sheet. I got 9 options per scene at once and chose from 55.
Video	Veo 3.1, Kling 3.0	Selected image + JSON prompt → 22 clips, 190 seconds of material
Dialogue and narration	Kling	Dialogue prompt with the speaker specified → shots and voice of the elderly speaker talking
Music	ElevenLabs	Prompt → background music
Editing	Premiere Pro	21 clips + voice → 140-second first cut
Finishing	After Effects	First cut → HUD overlay, upscale to 1920×1004, end credits

3×3 candidate sheet of an elderly person on a bench — Character sheet. Nine options, from the bench to a front view to an eye close-up.

3×3 candidate sheet of coastal rocks and tide pools — The “rocky shore” (돌이 많은 해안가) sheet.

3×3 candidate sheet of a harbor, fishing boats, and clasped hands — The “a place like a harbor” (포구 같은 곳) sheet.

Prompts as JSON

I wrote the video prompts in JSON, with a separate schema for each tool. For Veo, I split six elements into keys: cinematography, subject, action, context, style, and audio. For Kling, I put more weight on performance and speaker assignment, attaching each line directly to the person saying it. The criteria I wrote in my working notes:

The key is to follow the five elements in the Veo guide and structure cinematography + subject + action + context + style/mood + audio. (핵심은 Veo 가이드의 5요소에 맞춰서, 촬영 문법 + 피사체 + 행동 + 맥락 + 스타일/분위기 + 오디오를 구조화하는 것.)

According to the guide, its strength lies in being specific about camera, shot, narrative, and audio. In particular, linking dialogue directly to the speaker gives better alignment in multi-character scenes. (가이드상 강점은 카메라/샷/내러티브/오디오를 구체적으로 적는 것에 있다. 특히 대사와 화자를 직접 연결하면 멀티 캐릭터 장면에서 더 잘 맞춘다.)

The clips that remained in the final film can be told apart by their file signatures. Veo clips are all 8 seconds and leave a mark in the corner of the frame; Kling clips vary in length from 6 to 15 seconds and carry dialogue.

Source	Clips	Length	Use
Veo 3.1	14	8 s each	Landscapes, shots from behind, hand inserts
Kling 3.0	7	6–15 s	Shots of the elderly speaker talking, tablet scenes
Other	1	8 s	Fishing boat in the harbor

A grid of the first frames of the 22 clips used in the film — First frames of the 22 clips. I made 190 seconds of material and edited it to a 140-second first cut, then a 178-second final.

An elderly person seen from behind, sitting on a bench beyond a field of silver grass — Veo clip. The bench in the opening.

A tide pool between coastal rocks — Veo clip. The corrected memory: a rocky shore.

A high angle looking down on a sandy beach — The sea of the first memory. A high angle generated as a single image.

Editing and Finishing

In a little over two hours in the early morning of Apr 18, I made 22 clips and strung them together in Premiere Pro into a 2:21 cut. After sharing this first cut with the team, over the next five days I recorded five more dialogue shots and added four Kling clips. Finishing took five versions in After Effects. I placed HUD overlays on the right side of the frame — memory extraction detected, boats : 93%, scene reconstruction complete — to make visible how the platform pulls clues out of a memory. I upscaled the 1280×670 original to 1920×1004 and added slides on the process and expected impact, plus end credits. The final cut runs 2:58.

It did not pass the preliminary round. The video remains public as it is.