The ComfyUI team just published a behind-the-scenes look at how animator Xindi Zhang built an Oscar-shortlisted film entirely through node-based AI workflows. Her thesis, "The Song of Drifters," is a documentary animation about displacement and belonging, and it was made not with one-click generators but with hand-wired ComfyUI graphs that gave her control over every frame. The post, published June 23, 2026, doubles as a rare public teardown of a festival-grade AI pipeline.
What Happened
Zhang, who holds an MFA in Expanded Animation from USC and now works as an AI Creative at Amazon AI Studio, walked through the full production stack she used for her shortlisted thesis. Rather than prompting a model and accepting the output, she treated ComfyUI as a precision instrument, chaining nodes to enforce her own illustrative style across live-action footage, 3D scenes, and AI-generated transitions. Her guiding line: "It's not the tool that steals my art. It's the tool that expands my art."
Why It Matters
Most AI film coverage focuses on text-to-video models that produce a clip and stop. Zhang's workflow shows the opposite approach, where the creator stays in the driver's seat and the model serves a defined artistic vision. That distinction matters for anyone trying to make work that looks authored rather than generated, and it is increasingly what separates festival submissions from disposable clips. A node graph is reproducible, so once a look is dialed in it can be applied shot after shot without redrawing the prompt lottery each time. The same control-first philosophy underpins techniques like character consistency workflows, where consistency comes from the pipeline, not luck.
Key Details
Zhang's stack combined iPhone 15 Pro live-action footage and Blender 3D animation as base material, then applied style transfer built on Stable Diffusion 1.5 with IP-Adapter and ControlNet for precise control. She trained custom LoRAs on her own video footage of cities so the model learned her source material rather than other artists' work, and used AI morphing to blend abstract and concrete imagery for dream sequences. Both her style-transfer and morphing graphs are published openly in the Comfy-Org creative-campus repository.
What to Do Next
Clone the creative-campus workflows and load the style-transfer graph in ComfyUI to see how IP-Adapter and ControlNet are wired together. Start by feeding it a few seconds of your own phone footage and one reference illustration, then train a small LoRA on a folder of your own images so the style is genuinely yours. For editing the resulting sequences locally, pair it with our Bernini-R local video editing guide.
