“Dear Upstairs Neighbors”: Connie he draws on the noise from above

Anyone who has ever woken up at 2 a.m. trying to figure out what the people upstairs were doing will instantly recognize Ada. He’s tired, he wants to sleep, and the ceiling – along with his confused mind – has other plans.

What he does about it is where Dear upstairs neighbors It goes from a sitcom premise to something outlandish: a six-minute animated short in which a sleepless young woman’s guesswork on top of a rocket goes from high heels to skating octopuses in a perfect, death-metal-fueled battle for her sanity. It’s rendered in thick, dense, ever-moving brushstrokes, like an abstract expressionist canvas that learned to animate after drinking too much coffee.

Writer and director Connie Mann spent 2021 to 2024 Pixar As a story artist who works inside out 2 and Dream Productions; before his first short, Watermelon: A cautionary talefound an audience of tens of millions on YouTube. He and producer Markia Mayer, who voices Ada and now serves as head of production Google DeepMindmade Dear upstairs neighbors There with a 45 man crew of animators, technical artists and researchers. According to Cassidy Curtis and Sarah Rumble writes on the Google blogThe team built a video-to-video pipeline around custom Veo and Imagen models that are perfectly aligned to the production’s own hand-drawn concept art. The film had its world premiere at the Sundance Institute History Forum last January. Tribeca Festival in Whoopi’s Wonderful World of Animation shorts program in June.

Here’s the trailer:

The executive producers are Glenn Entis, co-founder of Pacific Data Images, who is now a senior director and executive producer of Google DeepMind; Douglas Eck, who directs generative media research there; and David Salesini, who leads the research team at Google. Yingzong Xin is the production designer. Curtis, who spent two decades at PDI/DreamWorks before moving to Google, is a supervising animator; Rumbley is the VFX supervisor; Ben Knight, Matthias Breitholz, and Stephen Chao are the character animators; Jung Spielberg composed the score, including the death metal song “Red Reverie” that sets the tone for the climax. Sarah Affleck replaced him. The research and technological side of credits includes several dozen names; The DeepMind post highlights researchers Andy Koenen, Forrester Cole, Ellen Jiang and Erica Lu.

It’s worth digging into the pipeline because “AI-assisted” covers many sins these days, often “DeepAmbiguity”; It’s not a normal workflow, though to be honest, Google has the resources to test, build, update like no other. Ada and most of her hallucinations were animated by hand in Maya and TV Paint. Generative models came last, in a position usually occupied by the renderer, to combine hand-crafted parts into artistic final frames. By the production’s own arithmetic, doing it by hand meant drawing more than 7,000 frames; Doing this in CG would normally have meant R&D of a feature film stylized render for a short of this length.

I emailed him a set of questions. Her answers follow, from the flat to the slider asking Ada’s face.

“The story started from my experience of living in an apartment building during the pandemic, where everyone was stuck in their rooms,” he says. “And I started hearing all kinds of noises—stilettos at 5 a.m., jumping in the living room several times a day, a six-month construction project—I had to leave eventually, but the noise didn’t stop when I moved to a new place. So one day I decided to sit in a cafe and put my frustrations on paper. Turning this story into a therapeutic experience.”

“After I had the first draft, I thought it would be a fun story and probably relatable to a lot of people. So I decided to keep tweaking the idea. From start to finish, it took my producer Marka Meyer and I 27 iterations to get to the final version of the story.”

He never thought of saying it any other way.

“Animation is a universal language,” he says. “And visual storytelling is my favorite form of storytelling because of how expressive it can be Dear upstairs neighbors The goal was to use visuals to convey feelings. Animation is the perfect playground for visual metaphor and the exploratory use of color and abstract texture.”

“I was inspired by both 2D and 3D animation. I love the flexibility of 2D animation and the cinematic quality of 3D animation, and I wanted to explore ways to bring these different visual languages together. Animation also gave us the freedom to really bring Ada’s imagination to life in a way that would have been difficult to achieve with Ada in live action. The all-encompassing power, The take they had on his mental state had to be surreal and surprising, but grounded in the way the story required.”

On where the look comes from and why death metal:

“The visual style of this film is largely inspired by my background, growing up in a family that loved art,” he explains. “My father is an abstract fine art artist, and I grew up in a lot of galleries and museums. I was always interested in the variety of artists’ self-expression and their voices through the use of colors, materials, or unique concepts. At the same time, as an animation enthusiast who is passionate about storytelling, I was always interested in what went with them. Ultimately, it inspired the art direction of this film.”

The death metal soundtrack was born out of my immense anger towards my neighbors. I couldn’t find a better type of music to evoke the emotion of anger than death metal. In fact, the look of the film’s climax is largely inspired by our Grammy-winning composer Jung Spielberg’s first demo. It had to be as interesting as the music became.”

A short series in three acts.

“It all started with a real experience when I was woken up by the noise of construction and drilling every day at 8am,” he says. “In the first act, everything came from Ada’s imagination with typical and ordinary sounds. We started with quite realistic sounds, like the sound of high heels, and gradually moved to more obscure sounds, octopuses and elephants jumping rope.”

“In the second act, we enter a more dream-like realm. I wanted to increase the intensity of Ada’s experience by having the objects ‘physically’ attack her body. By having the sound source and Ada in the same frame, the audience could also relate more to her experience.”

“In Act 3, Ada’s experience needed to be heightened and even more intense than in Act 2. I put her in an abstract, imaginative environment, and with a more cinematic camera language, the beautiful animation work of Cassidy Curtis, Stephen Chao and Ben Knight, and Marcia’s voice performance as Ada, we were able to capture those emotions perfectly.”

As for what viewers can expect from the film:

“We hope viewers have a fun ride on Ada’s emotional journey!”

The project moved from Pixar to Google DeepMind, bringing with it a new kind of production problem.

“Marcia and I were developing Dear upstairs neighbors About a year into Pixar, when a researcher who was familiar with our work introduced us to our executive producers, Glen Entis, David Salesin, and Douglas Eck at Google DeepMind,” he recalls. “The team’s researchers and engineers wanted to understand what artists needed and wanted in order to build tools that really supported the creative process. With their support, Marcia and I were able to bring to the team artists who were as eager as we were to explore responsible ways of using this new technology. The project became an opportunity to bring both parties together in the same room to learn from each other and, hopefully, create a model for the ethical and responsible use of generative artificial intelligence to improve the creative process.

“The biggest challenge was that this film was the first of its kind. There was no playbook and very few visual references. We had to build both the visual style and the production pipeline from scratch, figuring things out step by step. But with the support of our research and engineering teams, we were able to create innovative workflows that didn’t give that artistic vision. Straight into the artist’s eye.”

The models were refined on the production’s own concept art, and they took more than the crew expected.

“It was really exciting to see how the models were able to learn directly from the pieces we created,” he says. “For example, we gave the video model a set of concept art slides created by our production designer, Yingzong Xin. From there, we started noticing how the model reacted and interpreted those inputs.”

“Initially, we were mainly hoping he’d go with a painterly brush and a full color palette. But he actually went further, he also went with the low-angle, two-point perspective that we used all the time in the concept art.”

“We also added zig-zag brush strokes as a visual cue for sound effects when objects hit the ground. Interestingly, the model started to reflect this language as well, producing similar marks at these exact moments. This kind of emergent behavior is something you don’t usually see in traditional computer graphics workflows.”

The work process itself is divided into three parts. ControlNets, which he mentions below, allow the artist to guide the model with a pose or depth map rather than a text prompt.

“The work process Dear upstairs neighbors It can basically be divided into three main parts,” he says. “All pre-production was done by the artists, including storyboards, character designs, color keys and concept art. We then used this to teach the model our art direction by refining many custom models. We deliberately chose a style that is difficult to achieve with traditional computer graphics tools, but which also served well for visual storytelling.”

“Both the 2D and 3D animation was also created by the animators, which gave us precise control over performance, expression and timing. We also had renderers and modelers to create the characters and environments. From there, our R&D team built a bespoke pipeline to bring together the painterly visual style and animation. We implemented control meshes and mask-based image editing tools as for video.

What part of the final image did the artists control:

“We iterate a lot, almost the same amount you would on any production,” he says. “The artists had full control over all designs, animation choices, and on-screen pixels. The mask editing tool has settings from 0 to 1. 0 is no hallucination; 1 is a higher degree of hallucination, which in our case translates to a more artistic result based on the training set.”

“This gave our artists the ability to adjust different parts of the mask to achieve the look they were going for. For example, if we wanted Ada’s expression to read more clearly, we’d lower the strength on her face. But if we wanted her hair less CG and the background more defined, we’d increase the strength in those areas.”