Bokeh Dial — Depth-of-Field Control

Bokeh Dial — Depth-of-Field Control

A control-tool LoRA for Krea 2 that dials in camera-realistic shallow depth-of-field — soft, rounded background bokeh with a sharp in-focus subject — on ANY subject, any scene, any style. Think of it as a physical aperture knob you add to your prompt.

What it does

Unlike style LoRAs that lock in a look (era, color grade, lighting), Bokeh Dial changes only one optical property: how hard the background blurs relative to the subject. Apply it to a portrait, a product shot, a street scene, an animal, a landscape — the subject stays whatever style your prompt already describes, only the focus falloff changes.

Trigger word

bokehdial

Recommended strength

0.3–0.5 — Subtle background softness — editorial/documentary look

0.7–1.0 — Full studio-portrait bokeh (recommended default)

1.2–1.5 — Heavy, dreamy blur — telephoto/macro-lens feel

Strength scales linearly with LoRA weight — no special syntax needed. Negative weights weren't part of training but tested clean at -0.3 (sharpens the background slightly instead of blurring it) — experimental, use at your own discretion.

Example prompt

bokehdial, a chef plating a dish in a restaurant kitchen, indoor, night

How to see the effect

The bokeh isn't a style overlay — it's the depth-of-field itself, so a single still can look "just like a normal photo." What to look for: the subject (face, hands, product) stays crisp while everything behind it — signage, foliage, crowd, tools — dissolves into soft, rounded blur. It's most obvious at strength 1.0+, and clearest when you compare the same seed at strength 0 vs 1.0 — the falloff appears with nothing else in the image changing.

Training notes

• Base model: Krea 2 Turbo (fp8)

• Trained via: Civitai on-site trainer (AI Toolkit engine), 2000 steps / 10 checkpoints

• Dataset: 50 images, balanced portrait/landscape, diverse subjects (people, animals, objects, indoor/outdoor) — built so the effect generalizes instead of memorizing one scene type

• Showcase images: generated at strength 1.0, prompts and seeds outside the training set

Good to know

Works best when the prompt already implies a plane of focus (a subject in the foreground). Very text-heavy or signage-dense scenes can occasionally cause artifacting — if you see garbled text on surfaces, try a different seed.

Part of a 3-entry submission to the Krea 2 Training Contest — see also mossworld (World Morph) and y2ksnap (Art Style).