One game character, five style directions: planning a redraw matrix that survives review

A style direction is easy to argue about and hard to compare. Put three different characters in a mood board and you learn nothing about how one character reads under five lighting and material treatments — you have changed the subject, not the style.

This is a short working method for a game-character style matrix: fix the input, name the variables, reject identity drift before you pick a favourite. It comes from a guide I maintain on dlss5nvidia.com, where the sample is independent still-image style conversion, not a game runtime.

Fix the input first

Every direction has to start from the same source: same pose, same costume construction, same camera. Otherwise the second row of your matrix is a different character in a different shot, and the comparison collapses into taste.

Before writing the request, write down the identity anchors — head silhouette, exposed hair, coat opening, strap placement, equipment outline. Those are the things you will check after the redraw, not the things you hope the model respects.

In the pair we actually retrieved for the guide, the direction (rainy neon street, cinematic lighting) came back convincingly, and the identity did not: exposed face and hair became a helmeted head, and the framing shifted. That is a real result, saved with its request, and it fails the preservation brief. Worth stating plainly, because a plausible-looking image is exactly what makes reviewers skip the comparison.

Source: the site's own synthetic character sample, copied unchanged. Not a capture from a commercial game.

Result: an existing cached example from the same workflow. The helmet is an observed identity change, not an accepted conversion.

Keep the variables separate

Five directions are useful only if each one changes one thing you can see. Concretely, split the matrix by:

• Light: where the key light sits, and how much bounce colour reaches the face.
• Material: how cloth, metal and skin read under that light.
• Palette: which two or three hues carry the frame.
• Rendering density: how much fine detail the treatment adds, and where.

Mix all four in one prompt and no row teaches you anything — you cannot say which change caused the acceptability failure.

Reject drift before ranking

Order of review matters more than the ranking itself:

1. Identity anchors intact? If not, the row is a different concept, not a candidate.
2. Silhouette and costume readable at thumbnail size? Game visuals are read small and fast.
3. Text and interface meaning unchanged? If your frame has HUD text, health values, counters or subtitles, a generated approximation of lettering is a failure even when the rest looks coherent.
4. Only then: which direction reads best for the project.

Steps 1–3 are binary. Step 4 is the conversation.

What this method cannot tell you

A still-image matrix is a visual-direction discussion, not a rendering benchmark. It says nothing about frame time, how the look behaves in motion, or whether the design can be implemented in engine. Our examples are original synthetic and site-owned imagery, not captures from commercial games, and the guide separates each claim from what the saved files actually show.

If you want the full checklist, the saved prompt, and the two figures from the retrieved pair, the long version is here: One game character, five style directions.

---

I work on DLSS5-related visual tooling and research write-ups at dlss5nvidia.com; the site is independent of NVIDIA, AMD and Black Forest Labs.