Magnific co-founder Javi Lopez spent six months and 12 hours of generations on a solo AI backrooms film. Here's what he learned.
What is “Background Memories” and who made it? #
Background Memories is a 17-minute AI-generated short film set in the backrooms, produced entirely solo by Javi Lopez, co-founder of the AI image upscaler Magnific. Lopez took leave from his day-to-day role to work on it full time, expecting the project to take one or two months. It took six. The film relies on handheld-camera style AI video generations, layered sound design built in DaVinci Resolve, and a prompting process Lopez did entirely by hand, without agentic tooling or automation.
TL;DR #
-
Javi Lopez , co-founder of Magnific, spent roughly six months making a 17-minute AI short film set in the backrooms, working alone and full time on the project.
-
The final 17-minute cut represents about 12 hours of raw generations , since usable clips only came in short 15 to 30 second bursts.
-
Lopez wrote every prompt manually using a consistent structure covering setting, camera style, action, and dialogue, with no agentic workflows or prompt automation involved.
-
He described sound design as more important than the visuals themselves, layering generated music with library sound effects and original composition inside DaVinci Resolve.
-
His single biggest technical regret was choosing a handheld camera style , which made shots far harder to stitch together across generations than simple cuts would have been.
-
For his next backrooms episode, Lopez plans to build a small team (sound designer, voice actors, colorist) rather than repeat the solo process.
-
He recommends new long-form AI filmmakers use low-resolution draft generations to test prompts cheaply before spending credits on full-quality renders.
-
✕a coding agent
-
✕no-code
-
✕vibe coding
-
✕a faster Cursor
The one that tells the coding agents what to build.
How did one person make a 17-minute AI film? #
Lopez built the film shot by shot, generation by generation, without any pipeline to batch or automate the process. Every prompt followed a repeatable structure: he’d describe the setting and where characters were positioned, specify the handheld camera style, lay out the action, write any dialogue, and explicitly note “no music” so he could handle sound separately in post. There was no agentic system writing or iterating on prompts for him. It was manual, one clip at a time.
The scale of that manual effort becomes clear in the math. A 17-minute final film worked out to around 12 hours of total generated footage, because individual usable clips ran only 15 to 30 seconds. Getting a single handheld shot to work could take an average of 50 generation attempts, since the moving-camera style he chose is unforgiving when you try to splice clips from different generations together. A static or cut-based shot gives you more flexibility to mix and match good takes across multiple generations. A continuous handheld shot doesn’t forgive that kind of patchwork, so Lopez often needed one long, clean 30-second take rather than being able to stitch together fragments.
Why does sound matter more than people think in AI film? #
Lopez’s clearest technical lesson was about audio, not video. He said sound might matter even more than the picture itself for selling the feeling of a scene, and he didn’t treat it as an afterthought. He explicitly prompted his video generations with “no music,” then built the soundscape separately and deliberately in DaVinci Resolve.
His approach involved layering: taking an AI-generated song as a base, then stacking up to ten additional layers of sound on top, some generated with music tools, others sourced from sound effect libraries. That stacking is what gave the film its texture rather than the flat, slightly artificial feel you get when you just drop in a single AI-generated track under a video clip. For anyone doing long-form AI video work, this is a reusable insight: don’t rely on a video model’s native audio generation as your final sound. Treat it as one raw ingredient in a mix.
What was the biggest mistake in making the film? #
Lopez pointed to one specific decision as his key regret: choosing a handheld camera aesthetic. It fit the found-footage tone of the backrooms mythology, but it made the entire production dramatically harder. Because handheld shots can’t easily be assembled from mismatched generations the way cut-based editing can, he was stuck needing long, continuous, clean takes. That pushed his average attempts-per-shot up and extended the timeline well past his original estimate.
His advice for anyone starting long-form AI video work: do cuts. Standard editing, where you can pull the best few seconds from one generation and splice it against the best few seconds from another, is far more credit-efficient and far less painful than chasing a single unbroken take.
Is solo AI filmmaking worth it? #
Lopez’s answer, after finishing the project, was no, not again, at least not entirely alone. He admitted he was tempted multiple times to abandon the film before finishing it, and that working solo on something for six months wasn’t enjoyable even though the final result succeeded. For his next backrooms episode, he’s planning to bring on a small team: a professional sound designer, voice actors instead of generated voices, and a colorist, rather than doing every post-production pass himself.
This matters for anyone scaling up from short AI clips to long-form work. The tools make solo production technically possible, but the bottleneck isn’t the AI, it’s the human bandwidth to carry every creative decision (writing, prompting, editing, sound, color) across a project that takes months. Lopez’s experience suggests that even with capable generative tools, long-form AI filmmaking benefits from the same division of labor as traditional filmmaking.
What is draft mode and why does it matter for AI video budgets? #
One practical cost-saving technique Lopez highlighted is generating a low-resolution draft first. If a draft version follows the prompt correctly and matches the intended shot, only then do you spend the credits to generate or upscale it to full quality. This avoids burning expensive generation credits on prompts that need several rounds of tweaking before they land. For a project requiring an average of 50 attempts per handheld shot, that kind of cheap iteration loop is the difference between a feasible budget and a runaway one.
Frequently Asked Questions #
How long did Javi Lopez’s backrooms film take to make?
About six months, working full time, after initially expecting the project to take one to two months.
Did Javi Lopez use AI agents or automated prompting tools?
No. He wrote every prompt manually using a consistent structure (setting, camera, action, dialogue, music notes) with no agentic workflow involved.
What editing software was used for the film?
DaVinci Resolve, for both cutting the footage together and building the layered sound design.
What was the hardest part of making the film alone?
The handheld camera style he chose made it difficult to combine footage across multiple generations, often requiring around 50 attempts to get a usable continuous shot.
What is the backrooms, and why do AI filmmakers use it as a setting?
The backrooms originated as a single eerie image posted anonymously on 4chan in 2019, depicting an endless liminal space of empty rooms with unsettling entities inside. It grew into an open, shared internet mythology with no single owner, which makes it a popular and legally low-risk setting for independent and AI filmmakers to build original stories around.