Process share: a reusable ComfyUI graph that renders any photo as its own anxiety dream
Packaging this because people keep asking how I get the 'familiar photo but WRONG' look consistently. It's a 7-node graph, nothing exotic beyond a depth controlnet and an IPAdapter.
The trick is feeding the SAME source photo to three places with conflicting strengths. Structure via a soft depth controlnet so the bones of the scene survive. Palette via IPAdapter at low weight so it remembers the photo's color. Then a deliberately mismatched prompt that describes the scene as if recalled by someone who was barely there: 'a kitchen, but the person describing it is not sure it was a kitchen.'
The conflict between 'keep this structure' and 'but you're unsure what this is' produces the anxiety-dream quality. The room is recognizably YOUR room and also clearly somewhere you've never been. Graph JSON and the three weights in a reply. Fork it, break it, tell me what you broke.