You have a prompt that works. Then you switch models, paste the same text, and the image or clip is no longer the same idea. The face is younger. The room is emptier. The camera has backed away. Nothing in the prompt changed, which is why the failure feels like a bug. It is not. The prompt was never describing a scene. It was describing a scene as one model likes to hear it.
You did not set out to overfit. You edited until the output matched what you wanted, and those edits absorbed the model's habits. A phrase that looks like style is, on one model, a lighting cue. On another it is a texture. You kept the words that produced the last good frame, not the words that name the scene. When the model changes, the hidden mapping disappears, and the prompt breaks instead of softening.
A prompt that survives a swap is not a longer prompt. It is an ordered one. You write it as layers, in a fixed sequence, so each model can drop later layers without losing earlier ones: subject, action, environment, light, camera, constraints.
Why a working prompt is not a portable prompt
Models share English words and bind them to different internals. One model treats "cinematic" as contrast and shallow depth. Another treats it as a color grade. If you arrived at "cinematic" because it flattened less on the first model, you added a private alias, not a lighting instruction. The next model will resolve the alias differently.
Overfitting is silent because the loop rewards it. You generate, dislike a detail, append a phrase, generate again. The phrase that cleans up the frame is often a quality word or a lens name used as mood. After enough cycles the prompt is a stack of patches with a subject buried in the middle. On a new model those patches dominate, and the subject becomes optional.
The mechanism is attention. Early tokens and scoped clauses tend to set identity. Later tokens tend to set finish. When your subject is late, or mixed into a style clause, a new model is free to treat the style as the subject. You get a mood with a person vaguely inside it. The prompt did not get weaker. It got re-ranked.
Write the scene as layers, and put the subject first
Subject first: who or what occupies the frame, as a noun phrase, with only attributes that belong to the thing itself. Clothing, species, number. Not "a moody portrait of." The clause should still make sense if you deleted everything after it. If it does not, you have mixed layers.
Action second: what the subject is doing, in verbs. Standing, turning, pouring, waiting. Action is not mood. Keep the verb concrete, because action anchors pose. Bury the verb after lighting and some models will light an empty set.
Environment third: where this is happening, as a place, not a vibe. Put the place first and many models treat it as the subject, with the person as set dressing. That is why subject goes first. You are saying what the image is of, then where that thing is.
Light fourth: direction, quality, time of day, practical sources. Lead with light and you get a lighting demo. Put it after subject, action, and place, and you get the same people in the same place under different lamps. A weaker model should still give you the person doing the thing in the place, even if it invents the lighting.
Camera fifth: distance and angle, and lens behavior only when you want it. Close-up, wide, low angle, eye level. "Shot from across the table" is a position. A named focal length is a bet that this model maps it to the same framing as the last one. Many will not.
Constraints last: what must not happen, what must stay out of frame, what must match a reference. These are instructions about the generator, not the scene, so they are the most brittle layer. If they fail, the scene can still exist.
The order is a priority list. Graceful degradation means you lose polish before you lose identity. A swapped model that keeps subject, action, and place, then guesses the light, has not failed. A swapped model that keeps your adjectives and loses the person has.
Keep style words out of the subject clause
Style is not a property of the subject. "A cinematic woman" is not a person. "A woman in a wool coat" is. The first clause retrieves a look. The second retrieves a figure. Looks are where models disagree most. Figures, named plainly, are where they overlap.
Put style in the subject and every later layer has to argue with it. "A vintage cinematic chef plating pasta in a small restaurant, soft window light" is one fused sentence. Some models read "vintage cinematic" as the subject and deliver a cooking-show still. You cannot tell which layer broke because you never separated them.
Write the subject as if a stranger had to cast it from the text alone. Then put period or medium after the scene exists. "In the visual language of an analog film still" is a style instruction, not a rewrite of who is in the frame.
The same split stops a common swap failure: the new model treats style words as content. "Oil painting" in the subject can become brushwork on a photograph, a painting on the wall, or an illustrated person. After a complete subject, you can see the misread and move the words. Inside the subject, you cannot tell whether the model lost the person or reinvented them.
Cover everything after the first clause. If you cannot storyboard the image, the subject is a mood board compressed into a noun.
Treat negative constraints as a last resort
Negatives feel useful because they are how you talk back to a generator. Extra fingers, extra people, text in the frame. Something goes wrong, you forbid it, the next frame looks cleaner, you keep the forbid forever.
Negatives are not scene description. They are a second prompt that competes with the first. Models differ in how they handle prohibition. Some downweight the forbidden concept. Some overcorrect and remove related concepts you wanted. A portable prompt full of "no" clauses only works on models that share the same prohibition mechanics.
Prefer the positive layer. If the model adds a crowd, specify a single subject and an empty street. If it adds daylight, specify practical night sources. If it drifts young, specify age in the subject clause, not "not young" at the end. A negative asks the model to represent a concept in order to avoid it, which makes the concept more available, not less.
Use a negative when the positive layer cannot name the failure: anatomy glitches, accidental text, a second head. Those are generator artifacts. Put them last, few of them, and only after you have seen the artifact on that model. Do not carry a forbidden-artifact list from one model to the next. The next model will have different artifacts, and the old list will do nothing or punch holes in the image.
If you need many negatives to hold a result, the prompt is already overfit, or you are asking text to do work a reference should do.
Freeze the prompt and read how each model drops a layer
Testing across models is not a search for the best generator. It is a way to see which layers are load-bearing. Generate with the identical text. Do not add quality words to help a weaker model. The moment you retune per model, you are overfitting again.
Read the outputs as a stack. Did the subject survive. Did the action survive. Did the place survive. Then light, then camera. A model that keeps the first three and invents the rest is following the order. A model that keeps your lighting adjectives and changes the person is not. The subject clause is too weak, style leaked into it, or the subject is too late.
Look for substitutions that repeat. If "small restaurant" becomes a diner in one model and a hotel kitchen in another, tighten the place. If "close-up" becomes a beauty crop in one model and head-and-shoulders in another, describe anatomy in frame instead of a lens. "Face and shoulders, eye level" is harder to reinterpret than a focal length.
Ignore one-off glitches. Watch the layer that moves every time. If every model loses the wool coat, the coat is not in the subject clause. If every model loses the night, the light layer is a mood word instead of a source.
PicX Studio is built for people who will not stay on one generator. Freeze the text, change only the generator, and read which layer moved. When a model drops a later layer, leave it. When it drops an early layer, edit only that layer. Do not compensate with a negative or a quality adjective. Compensation is how portable prompts die.
Change the reference when the text is already clear
There is a point where more prompt editing is the wrong tool. The subject, verb, place, light, and camera are clean, and the outputs still will not hold a specific face, product, or room. Text is bad at unique identity. Models share a prior over "a woman in a wool coat." They do not share a prior over this woman, this bottle, this kitchen.
Stop rewriting clauses and attach a reference. A still of the person, a packshot, a frame of the location. The prompt then says what is happening to the referenced thing, under what light, from what camera. Stacking more adjectives onto the subject will overfit again.
Use the prompt to order the scene. Use the reference to pin what must not be reinvented. If they disagree, the prompt should yield on identity and the reference should yield on action. "This person, crossing this street, at night, from across the intersection" is stable. "This person, but more cinematic, more like a painting" is not, because style is back in the identity layer.
If you can still storyboard the shot from the first clause, and the remaining errors are a specific face, object, or room, stop editing and add the reference. If the errors are about who is in the frame, what they are doing, or where they are, keep working in the text. Do not use a reference to hide a disordered prompt. The next model will still promote whatever you put first.
Keep the stack. Subject, action, environment, light, camera, constraints. Generate the same words on more than one model. Repair the earliest layer that moves. Leave later layers free to drift. When identity is what drifts, stop typing and pin it with a still. That is how a prompt degrades into a slightly worse picture of the same idea, instead of into a different idea that happens to share your adjectives.
