First, the method
Change one thing per generation, the habit from Start Here. A fix you cannot attribute is a coincidence you cannot repeat. Diagnose, change the one word the diagnosis points at, and run it again.
The subject got ignored
You asked for a courier on a rooftop and got a rooftop. Cause: the subject drowned. Past roughly twenty words the model starts dropping the middle of the prompt, and adjectives are the first things that displace real content. Cut the subject back to one focal point, a verb, a count, and an anchor, as in Writing the Subject, and move style and light out of the subject into their own slots.
Twins, tiles, and second heads
Two identical subjects, a landscape that repeats itself, a torso that grew an extra head. Cause: the frame is too extreme for the content and the model filled the leftover space by repeating. The cure lives in the aspect ratio guide: pull the ratio back toward 3:2, or keep the wide frame and name what fills the sides so the model does not have to invent it.
Plastic skin and airbrushed faces
Portraits that look like rendered dolls. Cause: the model's default glamour finish, left unopposed. Oppose it with a physical medium: one film stock from the photography wall (Portra 400, Tri-X), film grain
, or a texture word. In Stable Diffusion, put airbrushed, waxy skin
in the negative prompt instead of piling realistic
into the positive.
Mangled hands
Six fingers, fused knuckles. Hands fail most when they are small, empty, and doing nothing. Three honest fixes: give the hands a job (holding a chipped mug
renders better than hands at rest), pose them out of frame (hands in pockets
, a cowboy shot), or fix the one bad hand afterward in a chat image model's edit loop, which the chat image models guide covers. The negative prompt (extra fingers, deformed hands
) helps in Stable Diffusion; it is a nudge, not a cure.
Muddy color, flat light
Everything is brown-grey soup, or evenly lit like a catalog. Cause: no palette instruction and no light direction, or two palettes fighting. Pick ONE word from the colors wall and one from the lighting wall that agree with each other. Golden hour wants warm palettes; neon noir wants teal and magenta. Two palette words in one prompt average into mud.
The set doesn't match
Ten images for one project and they look like ten projects. Cause: style words changed between generations, or were never stated and the model rolled a new default each time. Freeze one style word and one lighting word across the whole set, and for a recurring person use the frozen character block from Consistent Characters.
Garbled text
Signs and labels full of alien letters. Diffusion models draw the idea of text, not text. If the words matter, generate with a chat image model (Nano Banana and GPT-Image render short strings far better), keep the requested text under a few words and quoted, or design the image so the sign is decorative and add real type afterward.
When to stop fixing
Two targeted fixes that both failed means the problem is upstream: the subject phrasing itself is fighting you. Rewrite the subject from scratch in different words, or let the builder deal a fresh combination and carry your subject into it. By the fifth reroll of the same prompt you are playing a slot machine.
Published here first.