Prompting line · stop 02 of 8 · 24 min · members
Style, composition and identity are three different jobs. Using one reference for all three is why it fails.
Free with an account
Membership is free: an account opens all 86 script pages. The Lab, Studio Canvas and the paid guides need the $99 pass, paid once. Already signed in on this browser? The page opens by itself.
The problem
The model has to guess which part of it you meant, and it guesses badly.
Attaching a reference feels like the most direct instruction available: here, make it like this. In practice it is one of the vaguest, because a photograph contains a dozen separable qualities and the attachment carries no indication of which you cared about.
You wanted the light. The model matched the composition. You wanted the face. It matched the colour palette. The output is recognisably related to your reference and wrong in the way that matters, and no amount of re-rolling fixes it, because the ambiguity is in the request.
The fix is to know which of three distinct jobs you are asking for, and to attach accordingly.
The three jobs
Separable requests that people routinely bundle into one image.
These pull in opposite directions. A strong identity match constrains pose and lighting, which fights a style request. A strong composition match constrains where things sit, which fights identity if the pose differs.
Where your tool has separate slots for these, use them separately and set their weights independently. Where it does not, pick the one that matters most and express the other two in language.
Identity
A model cannot preserve a side of an object it has never been shown.
For identity, the useful reference is boring: even lighting, neutral pose, sharp, straight-on or three-quarter, nothing dramatic. Drama is style; it interferes.
The constraint people underestimate is coverage. If your reference shows the front of a product, a shot that turns it will invent the back. If it shows a face frontally, a profile will invent the side. The model is not retrieving a known object; it is extrapolating from one view.
So either plan shots that stay within what the reference covers, or supply multiple references covering the angles you need. Two clean views beat one beautiful one.
Do not use a graded frame
A heavily graded still carries its grade into everything downstream, and you will spend the rest of the project fighting a colour cast you supplied yourself. Use the ungraded original for identity references.
Style
If content and treatment both match, the model copies rather than applies.
For a style reference, choose an image whose subject is clearly different from what you are making. If you want the treatment of a car photograph applied to a bottle, use the car photograph — not a bottle photograph in that style, which invites the model to reproduce the reference bottle.
Keep the weight lower than feels natural. Style references at full strength tend to import composition too. Somewhere below half is usually where you get treatment without arrangement, though the specific number varies by tool and is worth testing once.
The most transferable qualities are palette, contrast, grain and light direction. The least transferable are anything that depends on the subject — a pose, a gesture, a specific lens effect tied to a particular distance.
Composition
Geometry transfers cleanly because it is the simplest thing to extract.
If a set of images needs to feel like a series, matching composition across them does more than matching style. Same framing, same subject placement, same margins — and the set reads as one shoot even when the treatment varies.
Composition references are also forgiving. A rough sketch, a blocked-out rectangle layout, or a photograph of something entirely unrelated will all transfer the arrangement, because the model is reading structure rather than content.
This is the practical route to a coherent set: one composition reference used across every frame, with subject and treatment carried by the prompt. It is far more robust than trying to hold style constant with words.
When not to use one
Two situations where attaching an image makes the work worse.
While exploring. A reference narrows the space immediately. If you do not yet know what the image is, that narrowing is premature — you will get variations on someone else's picture instead of finding your own.
When the prompt already works. Adding a reference to a prompt that was producing good results usually degrades them, because the reference and the language now compete. If the words are working, leave them alone.
The general rule: use a reference for what language cannot do — a specific identity, a precise geometry — and use language for everything it can. Most people do the opposite, describing identity in words and attaching a picture for mood, which is exactly backwards.
1:1 · two hours