Turn Existing AI Portraits Into a Complete Character Reference Set
Audit existing AI portraits, preserve useful references, and generate only the character views your set is missing.
The approved studio bust was already a valid frontal face reference, even though it was not 3:4.
Regenerating it would have solved the wrong problem. The character set did not need matching proportions. It needed complete reference coverage.
That distinction changes the entire review process.
When I look at existing character images, I no longer begin by asking whether they match one another as files. I ask what each image can still do. An approved portrait may not fit the final reference-sheet format, but it may still show the face clearly enough to remain useful. A scene image may not be suitable for identity, but it may still preserve the character’s in-scene appearance.
The task is not to make every asset uniform. The task is to give every approved image a clear reference role, identify what is missing, and generate only those missing pieces.
Start with an audit, not another generation
The first step happens before opening a generation tool.
Collect the approved character images and inspect them one at a time. Do not reject an image because a newer reference format has been chosen. Do not reject it because its aspect ratio differs from the rest of the set.
Instead, assign it a role.
The studio bust has a straightforward role. It can serve as the additional frontal face view. That remains true when its proportion is not 3:4. Its value comes from the identity information it preserves, not from the shape of its canvas.
This is the first acceptance test:
Does the image still provide a reference that the character set needs?
If the answer is yes, keep it.
This sounds obvious after the fact. It is less obvious when a clean new template makes older images look inconsistent. The temptation is to regenerate everything so the folder feels orderly. That creates more generation work without necessarily adding missing character information.
A neat set can still be incomplete. A mixed-format set can still be useful.
Separate reference function from file format
Aspect ratio is a delivery property. Reference function is an information property.
For the studio bust, the information is the frontal face. If the face is approved and useful, the image has passed its role test. Changing its proportion would not fill a missing view. It would only reproduce an accepted view in another container.
That is why I do not use matching aspect ratio as the acceptance test.
I use completeness.
The character set needs the references required to represent the character. The existing bust already covers one of those needs. The real gaps are usually elsewhere: the three-panel sheet and the full-body image.
Those are different assets with different jobs. Neither gap is repaired by remaking the bust.
This gives the audit a practical order. First, identify what the existing images already cover. Then compare that coverage with the set you still need. Only after that comparison should another image be generated.
Find the actual gaps
I usually find two pieces missing:
The three-panel sheet.
The full-body reference.
That is where I put the next generation effort.
The three-panel sheet contributes several character views in one asset. The full-body image adds information that a bust cannot provide. If those pieces are absent, the set is incomplete even if the portrait folder contains several strong faces.
This is where image count becomes misleading.
A folder can contain many approved portraits and still lack the required coverage. Several variations of the same bust do not become a full-body reference. A scene image does not become a neutral identity portrait merely because the character looks good in it.
The audit must be based on roles, not quantity.
I make the missing pieces explicit. If the frontal face already exists, I mark that role as covered. If the three-panel sheet does not exist, it becomes a generation task. If the full-body image does not exist, it becomes another generation task.
I do not add replacement tasks for approved assets that still perform their assigned roles.
Reclassify scene images honestly
A portrait with wardrobe and scene decoration can be useful without being a valid identity reference.
That image belongs in the set as the additional in-scene view. It shows the character within a scene context. That is a legitimate reference role.
It does not replace the neutral face.
This distinction matters because the setting and clothing do not provide the same kind of information as a neutral identity view. If the only close view of the character is tied to scene wardrobe and decoration, the identity portion of the set is still missing.
The correct move is not to discard the scene image. The correct move is to reclassify it.
Keep it as the in-scene reference. Then generate the neutral face separately.
This preserves the work that succeeded while addressing the exact gap that remains.
The wrong move would be either extreme. Treating the scene image as sufficient would leave the identity set incomplete. Throwing it away would remove a useful in-scene reference.
An audit avoids both errors.
Keep approved work unless it contradicts the new set
The preservation rule is simple: approved work is not thrown away.
It is completed.
There is one reason to move an older piece out of the active reference set. It contradicts the newer approved material.
A contradiction is different from a format mismatch. A non-3:4 bust is not automatically contradictory. It can still be the frontal face reference. A scene portrait is not contradictory merely because it includes wardrobe and decoration. It can still serve as the in-scene view.
The question is whether the old piece conflicts with the new character reference, not whether it looks different as a document.
If it does conflict, move it to _rejected/.
Do not delete it.
That folder records that the image no longer belongs in the active set while preserving the image itself. The asset is separated from production use without being destroyed.
This creates a clean decision boundary:
An image that still performs a role stays active.
An image that lacks a role but does not contradict anything can be reviewed without forcing a replacement.
An image that contradicts the new approved reference moves to _rejected/.
A missing role gets generated.
Use completeness as the final test
Once the missing pieces have been generated, review the set again by role.
The frontal face is covered by the approved studio bust.
The three-panel sheet is present.
The full-body image is present.
The in-scene image is classified as an in-scene reference.
A neutral face exists separately when the scene image cannot serve as identity.
Any older image that contradicts the new approved material has been moved to _rejected/ without being deleted.
That is what done looks like.
Notice what is absent from that test. There is no requirement that every image share the same aspect ratio. There is no requirement to regenerate an approved bust solely because it does not match 3:4. There is no requirement to discard a scene image because it cannot serve as the neutral identity reference.
The set passes when the necessary reference roles are covered without active contradictions.
What this costs
I never logged a monetary price or a fixed render count.
What I do have is a clear way to control unnecessary generation. Keep the approved bust. Keep the useful scene image. Generate the three-panel sheet, the full-body reference, and the separate neutral face when those pieces are missing.
The avoidable cost is regeneration that does not add a missing role.
Every time I am tempted to remake an image because its format looks out of place, I return to the same question: what information will the replacement add?
If the answer is none, I keep the approved image.
If the answer identifies a missing view, I generate that view directly.
Today, open one approved character folder and label each image by its actual role. Do not generate anything until you can name the missing reference.
A method is what still works on the morning inspiration doesn't show up.
