Realism Generator: one dial from illustration to photoreal

One design, one dial

The usual way to make a concept “more realistic” is to rewrite the prompt — and the design drifts with it: a new face, new costume details, a different pose. This generator keeps everything fixed — prompt, seed, base model, self-trained style LoRAs — and exposes exactly one control: realism strength. Turning it produces a ladder of the same character, from the painted game-illustration look toward photographic skin and materials.

Where it shows

The difference is clearest in the face and the materials: skin texture, light falloff on leather, the weight of the metal hardware. Silhouette, costume and pose stay where they were.

Calibrated per character

There is no single right value. Each character gets its own: around +0.7 where a modern outfit benefits from tactile materials, around +0.4 where an ornate costume or a stylised silhouette needs to stay readable. Past the sweet spot the backdrop starts to flatten and the image reads less like game art — so the ladder is rendered first and the value is picked from it, then fixed for the rest of the set.

The graph

Base model → self-trained style LoRAs (collapsed here) → the realism-strength node → prompt → sampler on a fixed seed. The realism node is an open community realism-slider LoRA; the pipeline around it — the style stack it rides on, the fixed-seed ladder and the per-character calibration — is the part built here.

How it is used

• Art-direction meetings — the same character at several realism levels, side by side, to choose the project’s tone.
• Keeping a series consistent — once a value is chosen, every new character in the set renders at the same level.