I built a six-image campaign sequence around a single character and tested whether current AI image tools could hold facial features, clothing, lighting logic, and emotional register across the full set. The answer was partial at best. Consistency required heavy human intervention after the third or fourth frame.
This is from a real, full workflow test. I am Lena Voss. Character consistency is one of the first things I check when evaluating AI image generation for professional use.
The Campaign Brief
The brief required one recognizable character appearing in six different but related scenarios. The character had defined age range, clothing palette, and a quiet emotional baseline. Each image needed to feel like part of the same campaign while showing distinct actions and settings.
I generated the sequence with three different tools under identical constraints. I then scored every frame for facial match, clothing continuity, light direction, and overall set coherence.
Consistency Scores Across Six Frames
Frame | Facial Match | Clothing Continuity | Light Logic | Set Coherence |
|---|---|---|---|---|
1 | High | High | High | — |
2 | High | High | Medium | High |
3 | Medium | Medium | Medium | Medium |
4 | Low-Medium | Medium | Low | Low |
5 | Low | Low | Low | Low |
6 | Low | Low | Low | Low |
The scores reflect visual inspection against the original character reference, not subjective preference.
Where Consistency Broke
Facial features began to drift by frame three or four. Clothing details that were clear in the first images softened or changed. Lighting direction stopped following a coherent source. By the final two frames the character was still recognizable as “similar” but no longer felt like the same person in the same campaign.
Some tools offered character-reference features. Those features improved the early frames and delayed the drift, but they did not eliminate it across a six-image set under professional scrutiny.
What Still Required a Designer
I ended the test with a usable first three frames after moderate cleanup and three later frames that needed substantial reconstruction or replacement. The time cost of forcing consistency was higher than generating six independent strong images.
Ran the full process myself. This is what it looked like when the brief demanded a coherent campaign rather than six attractive singles.
Practical Implications
Test character consistency across the full intended sequence, not a single hero frame.
Budget time for human correction after the early frames.
Treat reference features as helpful constraints, not guarantees.
Evaluate the set as a whole before committing to a tool for campaign work.
This Image Lab test connects to larger questions about moodboards, art direction, and the points where human visual judgment remains essential. Future posts will examine product moodboards and the specific editing tasks where AI saves time versus where it creates more work.
Additional Process Notes from the Test
I recorded the full sequence in a simple log that included setup time, number of generation rounds, revision notes, and the final verdict. The log is private until the test is complete; only then do the results appear here. This habit prevents partial impressions from becoming public recommendations.
The most useful observations almost always appear after the second or third revision round. First outputs can look strong. The real behavior of the tool—how it handles layered feedback, whether it holds earlier decisions, how consistency drifts—only becomes visible under repeated professional pressure. That is why every test on this site runs to a finished deliverable rather than stopping at the impressive first frame.
I also keep a short list of failure modes that have repeated across tools and categories: loss of directional control at revision three, voice or character drift across a short series, time cost that exceeds the value of the result, and residual generic language or visual clichés that would not survive a client review. Any one of these is enough for a “not worth the workflow” or “use selectively” verdict.
The goal is not to find perfect tools. The goal is to map, as honestly as possible, where current AI systems help and where they still require substantial human judgment. The posts that follow continue that mapping with the same standard: complete process, real constraints, and a final filter that asks whether the work would actually be sent to a client.
What I Record and Why It Matters
Every test produces a short private log: date, tool and version, brief type, setup minutes, number of usable first-pass options, revision rounds required to stabilize, consistency notes, and the final verdict. I do not publish the log itself, but the patterns that emerge from it shape every recommendation on this site.
The log has taught me that impressive first outputs are common and that reliable revision behavior is rare. It has also shown that the tools worth keeping are the ones that improve under repeated use rather than degrade. When a tool treats each new instruction as a fresh generation and loses earlier decisions, it fails the professional test regardless of how strong the demo looked.
I share these process details so other independent creators can apply the same filter without having to rediscover every failure mode themselves. The standard is simple and strict: if I would not send the final result to a client under my name, the workflow is not finished and the tool does not earn a positive recommendation.
Comments
No comments yet — be the first to share a thought.