Whitewash Cartoonization: Accidental Indexicality into Cartoon Space?

We present a technique for generating abstract, quasi-narrative animations through iterative image-to-image generation with lossy feedback transformations. The method feeds each generated image back into a diffusion model after applying a lossy transformation, creating sequences of arbitrary duration from a single prompt and optional input image. This technique emerged from an implementation error: a normalization bug causing brightness overexposure produced stylized, graphic outputs within several iterations, despite using generic models (Stable Diffusion Turbo) without stylistic prompts or inputs. We hypothesize the model interprets brightness-saturated images as drawings, where detail loss introduces abstraction and background homogenization. Similar effects occur with other lossy transformations including darkening, pixelization, and blurring. Beyond stylistic transformation, generated sequences exhibit emergent semantic dynamics—hands reaching, branches growing, figures moving closer and touching—despite these narrative elements being absent from prompts and the model lacking memory beyond the previous frame. We propose this resembles damage-induced regrowth processes like plant pruning or distraction osteogenesis, where information loss paradoxically stimulates regenerative responses. We have released open-source code and are developing a real-time system with variable controls for live visual performance and human-in-the-loop video editing through reactive modulation of noise and variation.