@autoaiai/aicut 0.1.11 → 0.1.13

package/skills/image-smixs/references/models.md

@@ -1,78 +0,0 @@
-# Model Selection — Nano Banana vs GPT Image 2.5
-
-Skill пишет промпты под две семьи моделей. Они мыслят по-разному — выбор модели меняет структуру промпта.
-
-## TL;DR
-
-| Задача | Модель |
-|--------|--------|
-| Реальное место/объект (с грунтингом) | **Nano Banana** (NB2/NBP) |
-| Сложная сцена с физикой/композицией | **Nano Banana Pro** |
-| Длинные горизонтальные/вертикальные форматы (1:8, 8:1, 4:1) | **Nano Banana** (только NB поддерживает экстрим) |
-| Дешёвая массовая генерация | **Nano Banana 2 Lite** или **GPT Image 2.5 Flare** (`quality: low`) |
-| Фотореализм с тонкой типографикой/UI | **GPT Image 2.5** |
-| Точное editing с preservation (try-on, swap, weather) | **GPT Image 2.5** (в editing у него лучшая identity-preservation) |
-| Маленький плотный текст в кадре | **GPT Image 2.5** (`quality: high`) |
-| Брендовая полиграфия / постеры с EXACT TEXT | **GPT Image 2.5** |
-| Сториборды, комиксы (последовательность) | **Nano Banana** (extreme ratios + thinking) |
-| Storyboard с фокусом на типографике | **GPT Image 2.5** |
-| Style transfer без упоминаемых референс-картинок | **GPT Image 2.5** (concrete visual targets) |
-| Рендер из 14+ референсов | **Nano Banana Pro** (до 14) или **GPT Image 2.5** (до 16) |
-
-> Внутри GPT Image 2.5 две модели: **Flare** (`gpt-image-2.5-flare`) — быстрый default, качество выше GPT Image 2 при ~50% меньшей латентности; **Sunburst** (`gpt-image-2.5-sunburst`) — базовая, для precision-edits и кампейн-качества. Промпт-правила одинаковые. GPT Image 2 — migration-only.
-
-## Когда что выигрывает
-
-### Nano Banana выигрывает в
-- **Image grounding.** NB2 ищет реальные изображения в интернете перед генерацией — точная архитектура конкретного храма, моста, площади; конкретные виды животных, растений. GPT Image 2.5 этого не делает.
-- **Экстремальные пропорции.** 1:8, 8:1, 1:4, 4:1 — баннеры, скроллы, комикс-стрипы. У GPT Image 2.5 max 3:1.
-- **«Thinking» режим.** Сложные инфографики со spatial logic.
-- **Цена/скорость.** NB2 = $0.04/img.
-
-### GPT Image 2.5 выигрывает в
-- **Identity preservation в edit.** Меняешь одежду / погоду / фон — лицо, поза, геометрия не плывут. Двухколоночная логика (change / preserve) работает как контракт.
-- **Тонкий текст в кадре.** Маленькие подписи, легенды, footnotes, multi-font layouts. На `quality: high` рендерит чётче.
-- **UI-моки и продуктовые скриншоты.** Иерархия, реальные интерфейс-элементы, читаемые лейблы.
-- **Структурированный 5-slot промпт.** Чёткое разделение Scene/Subject/Details/Use case/Constraints даёт предсказуемость.
-- **`quality` рычаг.** low/medium/high/xhigh/max — осознанный trade-off скорости и точности (шкала новая, пятиступенчатая; medium = default).
-- **4K-выход.** До 3840×2160 / 2160×3840 из коробки.
-
-### Где обе модели одинаково хороши
-- Photorealistic портреты.
-- Product shots на нейтральном фоне.
-- Минималистичные постеры.
-- Editorial-фотография.
-
-## Различия в синтаксисе промпта
-
-| Аспект | Nano Banana | GPT Image 2.5 |
-|--------|-------------|-------------|
-| Стиль промпта | Натуральный язык, 1-2 параграфа | 5-slot с лейблами секций |
-| Камера/линза | **Не указывать** числа (50mm, f/2.8) — NB игнорит | Можно «50mm feel», но как high-level look |
-| «Stunning/epic/masterpiece» | Игнорит, не вредит | **Anti-slop**: вредит, делает результат хуже |
-| Text in image | `"..."` в кавычках, font + position | `"..."` или ALL CAPS + «no extra words / no duplicate text» |
-| Negative framing | Использовать позитив | Использовать позитив + явный preserve list |
-| Сложные сцены | JSON для 5+ элементов | 5-slot template со секциями |
-| Edit | «Keep X same, change Y» | «Change: X / Preserve: Y / Constraints: Z» — повторять preserve каждую итерацию |
-| Множественные референсы | До 14, индексировать | До 16, индексировать с ролью («Image 1: base», «Image 2: jacket reference») |
-
-## Стоимость (ориентир)
-
-| Модель | Цена | Заметки |
-|--------|------|---------|
-| Nano Banana 2 Lite | ~$0.034/img | Только 1K, ~4 сек. Черновики и массовые батчи |
-| Nano Banana 2 (Flash) | ~$0.04/img | Default для большинства задач |
-| Nano Banana Pro | ~$0.15/img | Сложные сцены, до 14 рефов |
-| GPT Image 2.5 (`low`) | ~$0.01/img (1K) | Latency-sensitive, превью, массовые батчи |
-| GPT Image 2.5 (`medium`) | ~$0.024/img (1K) | Default для GPT Image |
-| GPT Image 2.5 (`high`) | ~$0.09/img (1K) | Маленький текст, brand-sensitive, photorealism |
-| GPT Image 2.5 (`xhigh`) | ~$0.16/img (1K) | Тонкие текстуры, print-ready |
-| GPT Image 2.5 (`max`) | ~$0.36/img (1K) | Максимальная fidelity, только по конкретной нужде |
-
-Цены Flare и Sunburst одинаковые (провайдерский ориентир на 1K; 2K/4K дороже). Каждый тир дешевле одноимённого тира GPT Image 2.
-
-> Скилл сам не запускает генерацию — выдаёт промпт. Модель/quality указываем рядом с промптом как мета.
-
----
-
-*Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*

package/skills/image-smixs/references/multi-panel.md

@@ -1,354 +0,0 @@
-# Multi-Panel Compositions
-
-Single-image layouts containing multiple frames, panels, or grid cells. Unlike [storyboards.md](storyboards.md) (sequential images generated one-at-a-time), these patterns produce **one image** with all panels baked in.
-
-Core principle: explicitly number and describe every panel. Models need panel-by-panel instructions — vague requests like "show multiple angles" produce inconsistent grids.
-
----
-
-## 1. 9-Cell Grid Storyboard
-
-> For a product-focused narrative variant, see [patterns/ecommerce.md](patterns/ecommerce.md#9-panel-tvc-storyboard-grid).
-
-**When to use:** TVC or commercial shot breakdown — one image = 9 panels with scene titles and timestamps.
-
-### Template (5-slot format)
-
-```
-Scene: Dark storyboard layout — 3x3 grid of cinematic frames for a {duration}-second {product_name} commercial. Each cell has a thin dark border. Below each cell: scene number, title, and timestamp in small white sans-serif text on dark background.
-
-Subject: {product_name} commercial storyboard showing the complete narrative arc across 9 panels.
-
-Important Details:
-Panel 1 (0:00–{t1}): {scene_1_description}
-Panel 2 ({t1}–{t2}): {scene_2_description}
-Panel 3 ({t2}–{t3}): {scene_3_description}
-Panel 4 ({t3}–{t4}): {scene_4_description}
-Panel 5 ({t4}–{t5}): {scene_5_description}
-Panel 6 ({t5}–{t6}): {scene_6_description}
-Panel 7 ({t6}–{t7}): {scene_7_description}
-Panel 8 ({t7}–{t8}): {scene_8_description}
-Panel 9 ({t8}–{duration}): {scene_9_description}
-
-Each panel labeled below: "Scene {N}: {title}" and "{start}–{end}"
-
-Use Case: Pre-production storyboard for video production team
-Constraints: All 9 panels must be clearly separated, no merged cells, every panel must contain distinct content, timestamps must be legible, {language} titles
-```
-
-**Variables:**
-- `{product_name}` — brand/product
-- `{duration}` — total spot length (e.g., "30", "60")
-- `{t1}` through `{t8}` — timestamp boundaries
-- `{scene_1_description}` through `{scene_9_description}` — shot description per panel
-- `{language}` — "Chinese" / "English" / "bilingual"
-
-**Recommended size:** 1536x1024 (landscape)
-**Model:** GPT Image 2.5 `quality: high` (text-heavy — timestamps and titles need legibility)
-**Common pitfalls:**
-- Omitting panel numbers causes the model to merge or skip panels
-- Vague scene descriptions produce near-identical panels — each must have a distinct action, angle, or subject
-- Timestamps in very small text need `quality: high` to remain readable
-
----
-
-## 2. 2x2 Editorial Portrait Grid
-
-**When to use:** Same person shown from 4 angles/crops in one image for an editorial or casting look.
-
-### Template (5-slot format)
-
-```
-Scene: 2x2 grid of four editorial portraits, minimal gap between panels. {background_description}. {lighting_style}.
-
-Subject: {subject_description} — same person in all four panels, consistent identity and wardrobe.
-
-Important Details:
-Top-left: front-facing portrait, direct eye contact, shoulders up
-Top-right: extreme macro close-up of face — eyes, skin texture, freckles visible
-Bottom-left: lower angle looking up, chin slightly raised, confident expression
-Bottom-right: side profile, clean silhouette against background
-
-Consistent lighting across all four panels. Skin texture realistic, no airbrushing.
-
-Use Case: editorial photography, model portfolio, casting composite
-Constraints: same person in every panel, no wardrobe changes between panels, no extra people, no text overlays
-```
-
-**Variables:**
-- `{subject_description}` — age, appearance, wardrobe (e.g., "woman in her 30s, dark curly hair, white linen shirt")
-- `{background_description}` — (e.g., "soft gray studio backdrop")
-- `{lighting_style}` — (e.g., "single key light from upper left, subtle fill from right")
-
-**Recommended size:** 1024x1024 (square) or 1024x1536 (vertical for portrait emphasis)
-**Model:** GPT Image 2.5 `quality: medium` or Nano Banana Pro (both handle photorealistic portraits well)
-**Common pitfalls:**
-- Not specifying "same person" in every panel — model may generate four different people
-- Omitting which quadrant gets which angle — model arranges arbitrarily
-- Requesting too-different crops (full body + extreme close-up) creates scale inconsistency within the grid
-
----
-
-## 3. 3-Panel Campaign Collage
-
-**When to use:** Hero shot + close-up + action for a campaign visual — horizontal or vertical triptych.
-
-### Template (5-slot format)
-
-```
-Scene: Three-panel {orientation} collage for {brand_name} campaign. Panels separated by thin white lines.
-
-Subject: {model_description} showcasing {product_description}.
-
-Important Details:
-Panel 1 (left/top): Hero wide shot — full figure of model with product, environmental context, lifestyle setting
-Panel 2 (center/middle): Close-up product detail — {product_detail_description}, tactile textures, sharp focus
-Panel 3 (right/bottom): Action shot — model using/wearing/interacting with product, candid energy, slight motion blur on extremities
-
-Consistent warm golden-hour lighting across all panels. Same model identity throughout.
-{typography_instruction}
-
-Use Case: social media campaign visual, brand lookbook
-Constraints: same person across all panels, consistent color grading, no stock-photo stiffness, {product_name} must be visible in every panel
-```
-
-**Variables:**
-- `{orientation}` — "horizontal" (side by side) or "vertical" (stacked)
-- `{brand_name}`, `{product_name}`, `{product_description}` — brand context
-- `{model_description}` — who appears
-- `{product_detail_description}` — what the close-up shows
-- `{typography_instruction}` — e.g., `Text: "{HEADLINE}" in bold condensed white, overlaid on Panel 1 lower third` or omit for no text
-
-**Recommended size:** 1536x1024 (horizontal triptych) or 1024x1536 (vertical triptych)
-**Model:** GPT Image 2.5 `quality: medium` — if text overlay needed, use `quality: high`
-**Common pitfalls:**
-- Not specifying panel order — "hero, close-up, action" without left/center/right assignment
-- Lighting inconsistency when mixing indoor close-up with outdoor hero shot — specify unified lighting
-- Forgetting to mention subject consistency causes three different models
-
----
-
-## 4. 4x3 Borderless Grid
-
-**When to use:** 12 panels telling a story or showing moods — no gaps between panels, seamless mosaic feel.
-
-### Template (5-slot format)
-
-```
-Scene: 4x3 borderless grid (4 columns, 3 rows) where each of the 12 panels is an independent image but panels share no borders, gaps, or dividers — edges touch seamlessly. Overall theme: {theme}.
-
-Subject: {subject_description} — maintain strong subject consistency across all 12 panels.
-
-Important Details:
-Row 1: {panel_1}, {panel_2}, {panel_3}, {panel_4}
-Row 2: {panel_5}, {panel_6}, {panel_7}, {panel_8}
-Row 3: {panel_9}, {panel_10}, {panel_11}, {panel_12}
-
-Style: {style_description}
-Mood progression: {mood_arc}
-Each panel is an independent composition but the overall grid reads as a unified artwork.
-
-Use Case: mood board, editorial spread, social media carousel preview, album artwork
-Constraints: borderless — no white lines, no black borders, no gaps between panels. Same subject identity across all panels. No text in panels.
-```
-
-**Variables:**
-- `{subject_description}` — who/what appears across panels
-- `{theme}` — overarching concept (e.g., "solitude in a city", "four seasons of a garden")
-- `{panel_1}` through `{panel_12}` — brief shot description per cell
-- `{style_description}` — visual style (e.g., "35mm film grain, desaturated palette")
-- `{mood_arc}` — how energy changes across the grid (e.g., "calm morning to chaotic night")
-
-**Recommended size:** 1536x1024 (landscape) — gives each cell enough resolution
-**Model:** Nano Banana Pro (handles complex multi-element compositions with thinking mode; no text needed)
-**Common pitfalls:**
-- Saying "borderless" is not enough — explicitly state "no white lines, no black borders, no gaps"
-- 12 panels in one image pushes detail limits — keep per-panel descriptions short and visually distinct
-- Without explicit subject consistency instructions, each panel may feature a different person/object
-
----
-
-## 5. 6-Frame Cinematic Sequence
-
-**When to use:** Fashion editorial or film-style sequence — multiple camera angles of the same scene in one image.
-
-### Template (5-slot format)
-
-```
-Scene: 6-frame cinematic sequence arranged in a 3x2 grid. Dark film-strip aesthetic with thin black borders. {location_description}.
-
-Subject: {subject_description}, wearing {wardrobe_description}. Same person, same outfit, same location across all 6 frames.
-
-Important Details:
-Frame 1 (top-left): Top-down bird's eye view — subject seen from directly above, full body visible against ground/floor
-Frame 2 (top-center): Low angle looking up — subject towering over camera, dramatic perspective, sky/ceiling visible
-Frame 3 (top-right): Wide isolation shot — subject small in frame, vast environment dominates, sense of scale
-Frame 4 (bottom-left): Close-up with slight tilt — face and upper body, Dutch angle, intimate intensity
-Frame 5 (bottom-center): Motion frame — subject mid-action ({motion_action}), natural motion blur on limbs
-Frame 6 (bottom-right): Grounded final — medium shot, subject at rest, direct gaze, resolving the sequence
-
-Photographer reference: {photographer_style}
-Lighting: consistent {lighting_description} across all frames
-
-Use Case: fashion editorial, film lookbook, director's shot list visualization
-Constraints: same person and wardrobe in every frame, no costume changes, consistent color grading, no text overlays
-```
-
-**Variables:**
-- `{subject_description}` — model details
-- `{wardrobe_description}` — clothing
-- `{location_description}` — setting
-- `{motion_action}` — what the motion frame captures (e.g., "walking forward", "turning sharply", "jumping")
-- `{photographer_style}` — (e.g., "Peter Lindbergh desaturated realism", "Helmut Newton dramatic contrast")
-- `{lighting_description}` — (e.g., "overcast natural light, soft shadows")
-
-**Recommended size:** 1536x1024 (landscape, 3x2 grid)
-**Model:** GPT Image 2.5 `quality: medium` or Nano Banana Pro
-**Common pitfalls:**
-- Not naming each frame explicitly — "various angles" is too vague, the model needs per-frame instructions
-- Top-down and low-angle in the same grid confuse the model if you don't anchor each frame to a grid position
-- Motion blur instruction must be specific ("blur on hands and feet") or the model applies blur everywhere
-
----
-
-## 6. Before/After Split
-
-**When to use:** Product transformation, makeover, time comparison, renovation — two states side by side.
-
-### Template (5-slot format)
-
-```
-Scene: Single image split into left and right halves with a {divider_style} dividing line down the center. Before/after comparison.
-
-Subject: {subject_description}
-
-Important Details:
-Left half (BEFORE): {before_description} — {before_condition}
-Right half (AFTER): {after_description} — {after_condition}
-
-The transition at the center line should feel {transition_style}. Same camera angle, same framing, same background perspective in both halves — only the subject's state changes.
-
-Use Case: {use_case}
-Constraints: identical composition and camera angle on both sides, same lighting direction, no text unless specified, the dividing line must be clearly visible
-```
-
-**Variables:**
-- `{subject_description}` — what is being compared
-- `{before_description}` / `{before_condition}` — left side state (e.g., "faded, cracked wall with peeling paint")
-- `{after_description}` / `{after_condition}` — right side state (e.g., "freshly painted wall, smooth finish, vibrant color")
-- `{divider_style}` — "thin white line" / "subtle gradient blend" / "sharp vertical cut"
-- `{transition_style}` — "clean and abrupt" / "natural, as if wiping away"
-- `{use_case}` — "product marketing", "renovation portfolio", "skincare results"
-
-**Recommended size:** 1536x1024 (landscape — gives each half a portrait-like proportion)
-**Model:** GPT Image 2.5 `quality: medium` — for text labels ("BEFORE" / "AFTER"), use `quality: high`
-**Common pitfalls:**
-- Not specifying "same camera angle both sides" — model may show two completely different viewpoints
-- Without a visible divider, the two halves can merge into one ambiguous scene
-- Left/right assignment matters — always state which side is before and which is after
-
----
-
-## 7. 12-Panel Storyboard Poster
-
-**When to use:** Full narrative in one image — 3x4 grid (3 columns, 4 rows) for animation or video pre-production.
-
-### Template (5-slot format)
-
-```
-Scene: 12-panel storyboard poster, 3 columns x 4 rows. Dark background with each panel in a clean rectangular frame. "{title}" in bold white text at the top of the image. Below each panel: scene number and one-line description in small white text.
-
-Subject: {character_description} — maintain consistent character design, proportions, and colors across all 12 panels.
-
-Important Details:
-Panel 1: {scene_1} — "{caption_1}"
-Panel 2: {scene_2} — "{caption_2}"
-Panel 3: {scene_3} — "{caption_3}"
-Panel 4: {scene_4} — "{caption_4}"
-Panel 5: {scene_5} — "{caption_5}"
-Panel 6: {scene_6} — "{caption_6}"
-Panel 7: {scene_7} — "{caption_7}"
-Panel 8: {scene_8} — "{caption_8}"
-Panel 9: {scene_9} — "{caption_9}"
-Panel 10: {scene_10} — "{caption_10}"
-Panel 11: {scene_11} — "{caption_11}"
-Panel 12: {scene_12} — "{caption_12}"
-
-Panels read left-to-right, top-to-bottom (like a comic page). Character appearance, clothing, and color palette must stay identical across all 12 panels — only pose, expression, angle, and environment change.
-
-Style: {art_style}
-
-Use Case: animation pre-production, pitch deck visualization, narrative overview poster
-Constraints: all 12 panels must be distinct scenes (no duplicates), character consistency is critical, scene numbers must be legible, {language} captions
-Quality: high
-```
-
-**Variables:**
-- `{title}` — project/episode title displayed at top
-- `{character_description}` — detailed character design (colors, outfit, distinguishing features)
-- `{scene_1}` through `{scene_12}` — visual description of each panel
-- `{caption_1}` through `{caption_12}` — text label under each panel
-- `{art_style}` — (e.g., "Pixar-style 3D", "anime cel shading", "watercolor illustration", "graphic novel ink")
-- `{language}` — caption language
-
-**Recommended size:** 1024x1536 (portrait — 3 columns x 4 rows needs vertical space)
-**Model:** GPT Image 2.5 `quality: high` (text-heavy — scene numbers and captions must be readable)
-**Common pitfalls:**
-- Character drift is the biggest risk at 12 panels — repeat character design details in the prompt, not just "same character"
-- Without explicit scene numbering in the prompt, panels may appear in random order
-- 12 panels with captions is extremely text-dense — keep captions under 5 words each for legibility
-- Art style must be stated once and applied uniformly; mixing styles across panels produces visual chaos
-
----
-
-## General Multi-Panel Tips
-
-**Panel count vs. detail trade-off:** More panels = less detail per panel. 4 panels allow rich per-panel descriptions; 12 panels need brief, visually distinct descriptions (3-8 words each).
-
-**Subject consistency:** Always include an explicit instruction: "same person / same character / same product across all panels." Repeat key identity markers (hair color, clothing, distinguishing features) rather than saying "same as before."
-
-**Grid specification:** Always state the grid dimensions (e.g., "3x2 grid, 3 columns 2 rows"). Saying "6 panels" without layout instruction lets the model choose an unpredictable arrangement.
-
-**Borders and gaps:** Be explicit — "thin white border between panels" or "borderless, no gaps." Default behavior varies by model and is unreliable.
-
-**Reading order:** State it: "left-to-right, top-to-bottom" or "numbered 1-9 starting top-left." Without this, narrative flow may be jumbled.
-
-**Panel variety:** State it as a requirement, not a hope: "no two non-empty panels may be identical or near-identical — each must change the shot size, the camera angle, the action beat, or the foreground/background relationship." Without this line the model repeats a flattering composition across half the grid. It is the mirror image of subject consistency: identity stays fixed, everything else has to move.
-
-**Model selection summary:**
-- Text/labels in panels --> GPT Image 2.5 `quality: high`
-- No text, complex composition --> Nano Banana Pro
-- Budget/exploration --> Nano Banana 2 or GPT Image 2.5 `quality: low`
-
----
-
-## Storyboard Sheets From a Timed Scene
-
-When the sheet feeds a video pipeline (see the sibling `video` skill), the scene's duration — not taste — decides the grid.
-
-**Grid from scene duration:**
-
-| Scene duration | Grid | Panels |
-|---|---|---:|
-| up to 9s | 3 columns x 3 rows | 9 |
-| 9-12s | 3 columns x 4 rows | 12 |
-| over 12s | 4 columns x 4 rows | 16 |
-
-Each panel stays 16:9 regardless of the grid. Keep the panels edge to edge — a thin dark rule is fine for readability, gutters are not.
-
-The duration in this table is the scene's screen time, not a clip length: the sheet shows the whole scene in one image, while the sibling `video` skill still generates it as several clips of about five seconds each. A 15s scene is one 4x4 sheet and three generations, not one.
-
-**Panel labels:** ask for two small solid-dark tabs inside every panel — the panel number in white in the top-left corner, the panel's own duration in white in the top-right ("0.5s", "1s", "2s"). Putting the duration on the panel instead of under it survives cropping and lets an editor read the sheet without the accompanying document. This is text — budget `quality: high` on GPT Image 2.5.
-
-**Default timing layouts:** start from an even split, then spend the extra time on the beat that needs a hold.
-
-- 5s on a 3x3 — eight panels at 0.5s, the last at 1s.
-- 9s on a 3x3 — nine panels at 1s.
-- 12s on a 3x4 — twelve panels at 1s.
-- 15s on a 4x4 — fourteen panels at 1s, the final two at 0.5s.
-
-Panel durations must add up to the scene duration exactly. If a short scene would fragment into meaningless beats, leave the trailing panels plain black rather than padding the action.
-
----
-
-*Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*

package/skills/image-smixs/references/nano-banana.md

@@ -1,150 +0,0 @@
-# Nano Banana — Specific Rules
-
-Три тира. «Thinking» модели — понимают намерение, физику, композицию.
-
-| Тир | Model ID | Роль |
-|-----|----------|------|
-| **NB2 Lite** | `gemini-3.1-flash-lite-image` | Самая быстрая и дешёвая ($0.034, ~4 сек), только 1K. Черновики, массовые батчи |
-| **NB2** (Flash) | `gemini-3.1-flash-image` | Дефолтный workhorse: Pro-фичи (текст, 4K, grounding, консистентность) на скорости Flash |
-| **NBP** (Pro) | `gemini-3-pro-image` | Топ-тир: сложные многослойные сцены, максимум reasoning и контроля |
-
-## Стиль промпта
-
-Натуральный язык, 1-2 параграфа. Структура свободная, но порядок помогает:
-**Subject + Action + Location/context + Composition + Style**
-
-```
-A cinematic wide shot of a futuristic sports car speeding through a rainy
-Tokyo street at night. Neon signs reflect off wet pavement and metallic
-chassis. Format: 16:9.
-```
-
-## Что НЕ указывать
-
-- Числовые параметры объектива: **50mm, 85mm, f/2.8, ISO 400** — NB игнорит. Используй описание: «shallow depth of field», «wide-angle distortion».
-- Tag-soup: «cool, modern, 4k, cinematic» — пишет связным предложением.
-
-## Уникальные возможности
-
-### Image Grounding (NB2 only)
-NB2 ищет реальные изображения в интернете до генерации. Архитектурно точные конкретные локации, корректные виды животных/растений.
-
-```
-Generate a cinematic, golden-hour photograph of [SPECIFIC REAL PLACE].
-Ensure the architectural details, the spire, the surrounding square, and
-the landscape are accurate to reality.
-```
-
-**Работает:** здания, мосты, площади, виды животных, виды растений, насекомые.
-**Не работает:** конкретные люди.
-
-### Extreme Aspect Ratios (NB2 only)
-
-1:8, 8:1, 1:4, 4:1 — для баннеров, скроллов, комикс-стрипов.
-
-```
-Create a 4-panel horizontal comic strip (aspect ratio 4:1).
-The story follows [CHARACTER] doing [ACTION] that ends with a twist.
-Use a vibrant comic book style. Keep character design consistent.
-```
-
-Стандартные: 1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9.
-
-### Thinking Mode
-
-Включён по умолчанию; у NBP отключить нельзя (модель рисует до 2 «thought images» в бэкенде, они не тарифицируются как картинки, но thinking-токены платные). У NB2 есть рычаг `thinking_level: minimal | high` (default `minimal`) — поднимай до `high` для:
-- Невнятных результатов, требующих reasoning
-- Сложных инфографик со spatial logic
-- Grounding + spatial reasoning одновременно
-
-### Multi-Reference (лимиты по тирам)
-
-| | NB2 Lite | NB2 | NBP |
-|---|---|---|---|
-| High-fidelity объекты | до 14 | до 10 | до 6 |
-| Character-consistency рефы | нет | до 4 | до 5 |
-| Style рефы | нет | нет | до 3 |
-
-```
-[Image 1: face]
-[Image 2: outfit]
-[Image 3: background]
-Combine: face from Image 1, outfit style from Image 2, setting from Image 3.
-Match lighting and perspective.
-```
-
-### JSON для сложных сцен (5+ элементов)
-
-```json
-{
-  "subject": {"description": "main subject", "expression": "emotion/pose"},
-  "photography": {"angle": "eye-level", "shot_type": "waist-up", "aspect_ratio": "16:9"},
-  "background": {"setting": "location", "lighting": "soft natural"}
-}
-```
-
-## Editing у Nano Banana
-
-Conversational, без масок:
-
-```
-Remove [OBJECT] from this image.
-Fill with [LOGICAL REPLACEMENT] matching surroundings.
-Keep [PRESERVED ELEMENTS] exactly the same.
-```
-
-После генерации добавляй итеративно:
-- «Make it warmer»
-- «Increase contrast»
-- «Soften background, add blur»
-- «Change headline color to #3b82f6»
-
-Interactions API держит контекст сессии — до 3 последовательных правок стэкаются без потери исходника. Правило «Edit, don't re-roll»: картинка верна на 80% — проси точечное изменение, а не новую генерацию.
-
-## Text Rendering
-
-- SOTA по 100+ языкам.
-- Font можно называть: «Century Gothic 12px», «Brush Script», «Impact», «Heavy blocky sans-serif».
-- Multi-language в одном кадре работает.
-- Hack: для сложного текста — сначала попроси модель **написать текст**, потом отдельным промптом «вставь этот текст в картинку».
-
-## Resolution / Cost
-
-| Resolution | Использование |
-|------------|---------------|
-| 0.5K | Самая дешёвая, batch и А/B варианты |
-| 1K | Default |
-| 2K | Финальный отбор |
-| 4K | Печать, hero-ассеты |
-
-**Workflow:** прогон вариантов на `0.5K` flash (или NB2 Lite на 1K) → отбор → переген победителя на `2K`/`4K`.
-
-В API `image_size` пишется строго с заглавной K (`1K`, `2K`, `4K`) — `1k` отклоняется. `512px` есть только у NB2; Lite — только 1K. Без указания размера модель матчит вход, иначе 1:1.
-
-## Кейфреймы для видео
-
-Nano Banana — image-слой перед motion-слоем. Всё критичное фиксируется на картинке ДО анимации: identity, композиция, aspect ratio, видимый текст, состояние продукта. Видеомодель усиливает любую неоднозначность стилла.
-
-- **Character sheet / hero frame.** Утверди один кадр-эталон персонажа, на него ссылаются все последующие video-промпты.
-- **Aspect ratio сразу целевой.** Кампании нужен вертикальный ролик — генери 9:16, а не кропай после анимации (тихая точка отказа).
-- **Хэндофф в видеомодель — motion brief, не пересказ.** Движение камеры, движение субъекта, движение фона, длительность, framing lock, запрещённые изменения. Видео-промпт не переописывает то, что уже есть в кадре.
-- Downstream любой: Veo 3.1, Kling 3.0, Seedance 2.0/2.5, Gemini Omni Flash (штатная пара Google, 10-сек клипы). Батч картинок → ревью-гейт → в видео уходят только одобренные стиллы.
-
-## Известные фейлы
-
-- Руки и лица всё ещё плывут (пальцы, суставы, дрейф likeness) — явные anatomy-инструкции + рефы + перегенерации. Есть отчёты, что NBP держит лица хуже NB2.
-- Мелкий текст мылится на 1K — для плотного текста 2K+ или GPT Image 2.
-- Инфографика может содержать фактически неверные данные — цифры проверять всегда, grounding помогает, но не гарантирует.
-- Склонность к overcooked HDR / пересату — проси «natural contrast, no HDR look».
-- SynthID-водяной знак стоит на всех генерациях (невидимый); на Vertex дополнительно C2PA.
-
-## Когда переключаться NB2 → NBP
-
-- NB2 не справляется со сложным многослойным промптом.
-- Нужен максимум референсов с разбивкой по ролям (6 объектов + 5 character + 3 style).
-- Photoreal с очень тонкой работой по материалам и свету.
-- NBP медленнее и дороже; на лицах иногда проигрывает NB2 — для портретной серии сначала тестируй оба.
-
----
-
-*Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*