I can't stop making cheap flash photos with MOSS-Image-Base (all prompts in the comments)
retained text
Ever since the realism settings went around here I've been stuck on one look: the harsh on-camera flash of a 2000s point-and-shoot. The background dropping into black, everyone a little shiny, somebody caught mid-sentence. MOSS-Image-Base gets that look really well if you describe the camera and what's wrong with the shot as well as the scene. I write the flaws into the prompt too: crooked framing, a little motion blur, red-eye on one friend, the fridge door hanging open a crack. My prompts ended up around a hundred words each, and a good part of that is the camera and the mistakes. My setup is MOSS-Image-Base bf16 with the realism graph everyone's been using: cfg 3, ER-SDE sampler with s\_noise 1.1, 25 steps on the simple scheduler, SLG on layers 22 and 24 for the whole run, negative left empty, roughly 2 megapixels. All stock ComfyUI nodes, no LoRAs, no upscaling. I ran three seeds per prompt and kept my favorite. All eleven prompts are in the first comment. The raccoon one still makes me laugh.
document fetched 2026-11-12 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xtl83k/i_cant_stop_making_cheap_flash_photos_with