Drift — A 6 Panel AI Generated Graphic Novel Page About a Memory That Never Happened
Built by DynamicMatrix ·
About this build
A single complete graphic novel page — 6 panels, full colour, consistent character design, coherent visual narrative — generated entirely with Midjourney and composed in Canva, telling a 90 second story about a woman who keeps remembering a conversation she is not sure ever took place.
What was built
The story the page tells: Panel 1 — Wide establishing shot. A woman in her late 30s sits at a kitchen table early morning. Coffee cup in both hands. Window behind her showing rain. She is looking at the empty chair across from her. Panel 2 — Close-up of her face. She is almost smiling. Remembering something. Panel 3 — A memory rendered in slightly warmer colour temperature. The same table. The same chair. Now occupied by someone whose face is turned away. Two coffee cups. Morning light instead of rain. Panel 4 — Back to the present. She reaches across the table toward the empty chair. Her hand stops halfway. Panel 5 — Her hand from above. Resting on the table. Not quite reaching. The second coffee cup is gone. Panel 6 — Wide shot matching Panel 1. She is still. The rain continues. The empty chair is empty. She is not looking at it anymore. No dialogue. No captions. Just 6 panels. The technical challenge — consistency: Sequential art lives or dies on consistency. The same character needs to look like the same person in every panel. The same kitchen needs to look like the same kitchen. The lighting needs to follow internal logic. AI image generation is notoriously bad at all of this because every generation is statistically independent — there is no memory between generations. The approach that worked: First built the character reference. Spent 3 days generating images of the protagonist until I had 4 images that felt like the same person from different angles and in different lighting. The final reference prompt was locked in with extreme specificity — 38 year old woman, dark auburn hair worn loose, high cheekbones, small scar above left eyebrow, wearing an oversized grey linen shirt, pale skin with warm undertones, tired eyes that still look like they are thinking about something. Then built the environment reference separately. The kitchen needed one specific quality — it needed to look like a place that had been lived in for years rather than a showroom. Scuffs on the table. A slightly chipped mug. A plant that had been there long enough to outgrow its original pot. Once both references were locked in used them as image prompts for every panel generation — always starting from the reference images rather than from text alone. Panel by panel production notes: Panel 1 and Panel 6 were generated together using the same prompt and the same seed value to maximise visual consistency between the bookend panels. The slight difference in her posture between the two is real — she sits slightly differently at the end of the page than she does at the start. Whether that reads as intentional or as AI inconsistency is something I genuinely cannot tell from the inside and would like feedback on. Panel 3 — the memory panel — was the most technically difficult. The warmer colour temperature needed to feel like memory without feeling like a filter had been applied. Generated 22 versions before finding one where the warmth felt internal to the light rather than external to the image. The turned-away figure in Panel 3 was a deliberate choice made after the first 8 attempts to generate a visible face resulted in faces that looked like different people in every generation. Turning the figure away solved the consistency problem and added an ambiguity that the story is better for. The person in the memory does not have a face. This might be the most honest thing about the page. Panel 5 — the hand from above — was the only panel generated without using the character reference. Hands in isolation consistently generate more naturally than hands attached to characters. The skin tone was matched to the reference in post using Canva's colour tools. The composition in Canva: Panel layout used a 2-3-1 grid — two panels across the top, three across the middle, one full-width at the bottom. The full-width bottom panel for Panel 6 was a late decision made after seeing all 6 panels together. The wide matching shot needed space to breathe. Added a very slight film grain overlay across all panels in Canva — approximately 8 percent opacity — to unify the visual texture across panels that had been generated at slightly different noise levels. This single step did more for the page's visual coherence than any individual panel decision. What worked better than expected: The emotional continuity. Three people who have seen the page without any explanation of the story have described the same emotional experience — something between melancholy and recognition. One described it as feeling like a memory they did not know they had. That is exactly what the page was trying to do and the fact that it landed without any words is the thing I am most surprised by. What did not work: Panel 2 close-up of her face is the weakest panel on the page. The almost-smile is there but it reads as neutral rather than as someone in the middle of remembering something. The face needs interiority and AI generated faces in 2026 still struggle with the very specific micro-expressions that signal internal experience rather than emotional state.