Style Variant Generator · öffentlicher Open Skill
Style Variant Generator
Style Variant Generator ist ein öffentlicher Open Skill auf Digen. Öffne ihn, um zu sehen, was dieser Workflow erstellt.
Beauty
52 6 14
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3954846/2026-09-11/azgh6fek.png
# Style Variant Generator ## When to use - The user uploads a reference image or gives text as the **subject/content**, and wants 4 images of different poses / outfits / looks generated in the **fixed** art style. - The user wants "generate 4 images of different looks in the same art style" — 4 style-consistent variants from a subject image or a text description. - The user's input is only the **content/subject**; the art style is always the fixed one below and must NOT change based on the user's image or text. ## Fixed art style (ALWAYS use this — never change it) The style is anchored by the 6 preset reference images `style_ref_1` … `style_ref_6`. It is a **direct-flash, film-grain, theatrical avant-garde editorial photography** look, spanning two related moods: - **90s Japanese street fashion editorial** (anchored by `style_ref_1`, `style_ref_2`, `style_ref_3`): analog film snapshot aesthetic, direct on-camera flash with hard frontal lighting, rapid light falloff into darkness, night documentary atmosphere, desaturated warm earth tones (cream / beige / oatmeal / ecru), muted film grain, theatrical androgynous avant-garde fashion styling (oversized suits, utilitarian couture, harness straps, chunky boots), often with surreal/yokai touches (ram horns, chalk-white face paint, silver hair). - **Surreal luxury noir / mob-boss** (anchored by `style_ref_4`, `style_ref_5`, `style_ref_6`): direct on-camera flash, low-key moody chiaroscuro, desaturated warm vintage tones (gold / ivory / black), film grain, opulent decadent wealth signifiers (cash, champagne, gold jewelry, marble), surreal creature/anthro figures (elephant-headed humanoid), noir "old money" atmosphere. - Shared across all 6: **photographic realism** (no painterly, anime, or 3D-render look), direct flash, desaturated warm film tones, visible grain, night/decadent atmosphere, theatrical surreal avant-garde styling. This style is **fixed**. Whatever image or text the user provides is only the subject/content; the visual style must always match the 6 preset reference images. ## Steps 1. **Determine the subject/content** from the user's input (this is NOT the style): - If the user uploaded a subject image: use `recognize_image` on it to describe the **subject** (what/who it depicts, pose, outfit, expression) — do NOT adopt its art style. **Describe the character's identity in detail** (face shape, facial features, hair style/color, build, age, and any distinguishing traits) so it can be preserved consistently across all variants. - If the user gave text: use the text as the subject description. - Combine both if present. 2. **Choose the 4 variant directions from the reference images** (poses / outfits / scenes / expressions). Do NOT always reuse the same fixed set (e.g. standing full-body / seated half-body / walking / close-up). Instead, look at the poses, outfits, and scenes shown in the preset reference images — especially `style_ref_4`, `style_ref_5`, `style_ref_6` (e.g. seated at a marble desk counting cash, lounging in a limo raising a champagne flute, sprawled on a leather banquette with an entourage) — and pick 4 distinct directions drawn from those reference poses/outfits/scenes. Vary the selection each run so the results are not the same four every time. If the user specified directions, use theirs instead. 3. **Establish a character anchor** (only when a subject image is provided — this is the key to strong facial consistency). Do NOT feed the subject image together with all the style refs in one shot, because the style refs dilute the character's identity. Instead: - Generate one clean **front-facing portrait** of the subject in the fixed style, using the subject image as the **primary character reference** and only 1 style reference image (`style_ref_1`) for the look. Use `model='i2i.hd'`, `input_urls` = [subject image, `style_ref_1`]. Name it `character_anchor`, `entity_type` = `style_variants`. The prompt must state the fixed style + the c