Partition Summer
A comic memoir of the 1947 Partition of India, as seen through my grandfather’s childhood memories
Partition Summer is a short comic I created based on stories my grandfather told me about living through the Partition of India in 1947. He was six years old at the time.
Behind the scenes
In early 2025, I spent around an hour every morning for about a month video calling with my grandfather (who I call “Pappi”). It started as a way for me to practice Hindi before my son was born while getting to know Pappi’s life story. I took detailed notes – the stories he told about being a six year old child during Partition were wild. They seemed like they were straight out of a movie. I had no plan for what I would do with the notes.
My son was born a few months later and I became busy with taking care of him and other projects, but every so often I’d brainstorm on what I could make with my notes. I wanted something that my son could read when he was older and Pappi’s stories reminded me of the graphic novel Maus, so I settled on making a short comic. Unfortunately my artistic talent is close to zero, so it would have taken a very long time to learn how to draw/illustrate before being able to produce something I’d be satisfied with. I played around with LLM image models for this, but this was late 2025 & they couldn’t yet do a great job without a ton of effort on my end. Finally in early May 2026, the image models (I used ChatGPT primarily for the comic) got good enough to produce output I was happy with.
Coincidentally my grandfather’s 85th birthday was on June 4 2026, so my mom suggested I finish the comic as his birthday present. I spent a few weeks making the comic, my uncle in India got it printed, and we gave it to him on his birthday!
Reflections on the process
I structured this as a project in ChatGPT and had a few key things in the sources:
- My detailed notes – source of truth for many decisions in the comic
- A page by page detailed storyboard for the comic that I had iterated on first with an LLM before starting any image generation – I ended up deviating a decent amount from this, but it was key for the first pass
- A historical account I found online (The Qadian Diary) – key for corroborating a few historical facts & for a map of what Qadian looked like in 1947
- A file with the number & ages of all the siblings at each point in the comic (1947 and 1952)
Before starting any image generation, I also discussed an overall comic co-creation workflow with ChatGPT and used the result as the project instructions. I explicitly asked ChatGPT to take into account known limitations of image generation models while deciding on this workflow. I also settled on a simple approach to version control for the comic.
The absolute most annoying thing about this whole process was keeping things consistent. I think my upfront work was quite helpful, but there’s more I would do next time. Specifically:
- Locking the aspect ratio/borders/margin of each page before doing anything. Fixing this at the very end was horrendous
- Adding the family continuity lock with the ages of all the siblings, role of the mother/father, etc much earlier – I only added it to the project source when I was close to done
- Adding character sheets – regenerating pages to add in siblings or change what age they looked like was nearly impossible without affecting the art
- Making the captions an overlay layer that is separate from the art
Other miscellaneous notes:
- I settled on pretty much the first art style the LLM generated; next time I would spend much more time exploring the design space of what the art could look like
- LLM image gen models do way better with positive instructions (“change this”) rather than negatives (“this doesn’t look right”)
- Providing nearby pages as context for specific characters/style when editing a page is surprisingly effective
- Asking the LLM to give me an image gen prompt for how to fix a particular issue is a useful first pass
- LLMs can’t handle more than a single page at a time
- I tried to make this comic a couple times and thought the image gen was shit each time – great lesson that in an era of LLMs it’s sometimes better to just wait until it’s easy lol