[style] Style-transfer branch — board vs pack vs hero image (IPAdapter Plus/Style)

[taskmasterpeace]

Sub-issue of #88. Build the STYLE branch (independent of identity) via IPAdapter Plus/Style. Test Robert's 'style-die-sheet' idea 4 ways on the SAME scene: (1) text-only style prompt, (2) one hero style image, (3) the 6-panel style board (horse/underwater/beach/apple/living-room/skyline), (4) 3 separate style images as a batch. Style ref = the comic still (red-dress woman + cowboy, mall). Clean 2×3 board, no labels/borders/text. Start at MODERATE style strength (too high = copies a panel's composition/palette; too low = generic FLUX/SDXL). Score: style match across person+horse+city+water, and whether it copied a panel's layout. Hypothesis: 3-image batch beats the collage.

[taskmasterpeace]

Direction locked — Krea 2 native Style Reference (the "board" model)

Going with Krea 2's native Style Reference / moodboard, not IPAdapter:

• Up to 4 weighted style-reference images, blended — each with its own strength % (0–100, default 50). Low % = subtle/realistic, high % = full style (palette + brush strokes). Combining weights mixes styles creatively without touching the prompt.
• UI: drag images into an "add as style transfer" zone next to the prompt; each ref is a thumbnail with a centered ▲ / NN% / ▼ strength stepper + a × remove — no separate slider/meter.
• Plumbing: cupcake via mk-gateway POST /v1/image with style_refs: [{upload_id, weight}] (each image staged through /v1/uploads). Same endpoint serves both Palette and DD.

Being built for Palette's Shot Creator first (Krea Phase 2); the box graph is in progress. Re: board vs pack vs hero — this is the board (multi-ref weighted blend).

[taskmasterpeace]

Update: Palette Phase 2 shipped → #224 (directors-palette-v2). The style-transfer UI is built — drag up to 3 refs, each a thumbnail with a ▲/%/▼ strength + ×, wired to mk-gateway /v1/image style_refs. Box graph live-verified (impressionist + pop-art transfer, 2-ref blend; box commit 10f3cc0). Local-turbo limits to note for the DD UI too: cap 3 refs (Qwen edit encoder has 3 image slots) and a single averaged global strength (true per-image strength would need Krea's paid cloud partner node). DD reuses the same /v1/image style_refs endpoint.