vargai 0.3.2 → 0.4.0-alpha2

package/docs/prompting.md

@@ -0,0 +1,326 @@
+# Prompting Kling & Nano-Banana-Pro
+
+_Shared by Alex_
+
+---
+
+## Kling AI (Video Generation)
+
+### Core Prompt Structure
+
+A strong Kling prompt relies on four key elements: **subject**, **context**, **action**, and **style**. If you omit any of these, the model is forced to guess.
+
+**Formula:**
+
+```
+[Subject] + [Action/Motion] + [Context/Setting] + [Camera/Style]
+```
+
+### What Makes Kling Special
+
+- Kling's greatest strength is **camera motion** and **character physics**
+- Works best with prompts under **40–50 words**, using clear and structured formatting
+- The real magic shows up when you guide it visually, not just verbally—**reference-first prompting** with images
+
+### Pro Techniques
+
+#### Camera Control
+
+- Add detailed camera movements: `"slow zoom-in"`, `"quick pan"`, `"aerial shot"`
+- Include technical specifications like `"Shot on virtual anamorphic lens, 24mm, f/2.8"`—these function as stylistic cues
+
+#### Weight Elements
+
+Use emphasis indicators `(++)` for critical elements:
+
+```
+++sleek red convertible++ driving along coastal highway
+```
+
+#### Negative Prompts
+
+Specify what to avoid:
+
+```
+No people, no text overlays, no distortion in vehicle proportions
+```
+
+### Image-to-Video Formula
+
+For Image-to-Video, the essential elements are **subject** and **movement**. Unlike Text-to-Video, which requires scene description, Image-to-Video already has a scene provided by the input image.
+
+**Example:**
+
+```
+Cat walking forward on the alien landscape, his tail swaying gently. 
+Vibrant meteor shower fills the sky, with meteors streaking across.
+```
+
+### Common Pitfalls
+
+- Requesting `"360-degree rotation around subject while zooming in"` often produces warped geometry due to multiple simultaneous camera transformations
+- Avoid relying on specific numbers—the AI may struggle with consistency (e.g., "5 trees", "6 puppies")
+- Mixing lighting terms like `"golden hour"` with `"studio lighting"` confuses the model's style interpretation
+
+---
+
+## Nano-Banana-Pro (Image Generation)
+
+Google's latest model—internally "Gemini 3 Pro Image."
+
+### Key Paradigm Shift
+
+Nano-Banana Pro is a **"Thinking" model**. It doesn't just match keywords; it understands intent, physics, and composition. Stop using "tag soups" (e.g., `dog, park, 4k, realistic`) and start acting like a **Creative Director**.
+
+### Core Prompt Formula
+
+```
+[Subject + Adjectives] doing [Action] in [Location/Context]. 
+[Composition/Camera Angle]. [Lighting/Atmosphere]. [Style/Media]. 
+[Specific Constraint/Text].
+```
+
+**ICS Method:** Always specify:
+
+1. **Image type** — blueprint, infographic, diagram
+2. **Content** — source data or information
+3. **Visual style** — survival guide, McKinsey presentation, comic
+
+### Killer Features
+
+#### Text Rendering
+
+Don't just say "add text." Be specific:
+
+```
+Write the text 'HELLO WORLD' in a bold, red, serif font on the sign.
+```
+
+Maximize text legibility by isolating string literals in double quotes. Explicitly define font family.
+
+#### Conversational Editing
+
+The model excels at understanding conversational edits. If an image is 80% correct, don't regenerate from scratch—simply ask for the specific change you need.
+
+#### Search-Powered Generation
+
+Nano-Banana Pro uses Google Search to generate imagery based on real-time data, current events, or factual verification.
+
+#### Multi-Image Context
+
+Supports a **14-image context window**. Upload images (style guides, logos, character sheets) and instruct:
+
+```
+Use the uploaded images as a strict style reference...
+```
+
+### Pro Tips
+
+- Remove polite phrases like "please"—use command-line style syntax
+- Trick the model into photorealism by specifying camera gear: `"Shot on Arri Alexa"` forces the AI to emulate specific film grain
+- You don't need `"4k, trending on artstation, masterpiece"` spam anymore—Nano Banana Pro understands natural language
+
+### Limitations
+
+- Rendering small text, fine details, and producing accurate spellings may not work perfectly
+- Always verify the factual accuracy of data-driven visuals like diagrams and infographics
+
+---
+
+## Quick Comparison
+
+| Aspect | Kling | Nano-Banana-Pro |
+|--------|-------|-----------------|
+| **Type** | Video gen | Image gen |
+| **Sweet Spot** | Camera motion, physics | Text rendering, editing |
+| **Prompting Style** | Cinematic director | Creative director |
+| **Best Input** | Image + motion prompt | Natural language + refs |
+| **Iteration** | Re-generate | Conversational edits |
+
+---
+
+## TikTok-Style Prompts for Kling
+
+### General Guidelines
+
+- **Image-to-Video** gives more control (upload frame → describe movement)
+- Specify **vertical 9:16** aspect ratio
+- Keep prompts short (**40–50 words max**)
+- Use **one camera movement type** at a time (don't mix pan + zoom + rotate)
+
+### Timeline Breakdown
+
+| Time | Action/Angle | Notes |
+|------|--------------|-------|
+| 0–1s | Hook — unusual expression or text question, frontal close-up | Capture attention in first 1–2 seconds |
+| 1–4s | Switch to 45° medium shot; first grimace | Sharp angle change on music beat |
+| 4–7s | Low angle or extreme close-up — funny angle; add text/emoji | Use jump cut to remove pause |
+| 7–10s | High angle; new emotion, possible match cut or object cover | Insert reaction cutaway for smoothness |
+| 10–13s | Show all four angles in 2×2 grid (optional) | Creates powerful final effect |
+| 13–15s | Finale and CTA: return to frontal shot; deliver joke or call-to-action | End on strong beat; add text with CTA |
+
+---
+
+### Prompt Examples by Segment
+
+#### 0–1 sec: Hook — Frontal Close-up
+
+**Text-to-Video:**
+
+```
+Young woman, surprised expression with wide eyes, looking directly at camera. 
+Extreme close-up face shot, static camera, soft ring light, vertical 9:16 format.
+Cinematic, social media aesthetic, sharp focus on eyes.
+```
+
+**Image-to-Video:**
+
+```
+Woman's eyes widen in surprise, eyebrows raise slightly, subtle head tilt forward.
+Static shot, no camera movement.
+```
+
+**Negative prompt:** `blur, camera shake, side profile, horizontal format`
+
+---
+
+#### 1–4 sec: 45° Medium Shot + Grimace
+
+**Text-to-Video:**
+
+```
+Woman at 45-degree angle, medium shot showing torso and hands. 
+She makes an exaggerated funny face, hands gesture expressively.
+Camera slowly pushes in, soft natural lighting from window, vertical format.
+TikTok influencer style, playful mood.
+```
+
+**Image-to-Video:**
+
+```
+Woman turns head 45 degrees, makes exaggerated grimace, 
+hands move up near face expressively. Slow push-in camera movement.
+```
+
+**Tip:** Use `++grimace++` or `++expressive hands++` for emphasis
+
+---
+
+#### 4–7 sec: Low Angle or Extreme Close-up
+
+**Option A — Low Angle:**
+
+```
+Woman filmed from below, low angle shot, looking down at camera with 
+confident smirk. Dramatic perspective, she appears powerful and playful.
+Static camera, vertical 9:16, slight lens distortion for comic effect.
+```
+
+**Option B — Extreme Close-up (eyes/lips):**
+
+```
+Extreme close-up of woman's eyes, pupils dilate slightly, 
+eyebrow raises in surprise. Macro shot, shallow depth of field, 
+ring light reflection in eyes. Static, vertical format.
+```
+
+**Negative prompt:** `full body, wide shot, horizontal, distortion in face proportions`
+
+---
+
+#### 7–10 sec: High Angle + New Emotion
+
+**Text-to-Video:**
+
+```
+Woman filmed from above, high angle shot, she looks up at camera 
+with playful vulnerability. Arms spread wide filling the frame.
+Slow gentle tilt down, soft overhead lighting, vertical 9:16.
+Cute, endearing mood, TikTok aesthetic.
+```
+
+**Image-to-Video:**
+
+```
+Woman looks up at camera, expression shifts from neutral to excited smile,
+hands move into frame from sides. Subtle camera tilt, gentle movement.
+```
+
+---
+
+#### 10–13 sec: Split Screen 2×2 (All 4 Angles)
+
+Kling doesn't natively create grids—this is post-production. Options:
+
+1. **Generate each angle separately** → combine in CapCut/Premiere
+2. **Experimental mosaic prompt:**
+
+```
+Split screen 2x2 grid showing same woman from four angles simultaneously:
+top-left frontal close-up, top-right 45-degree medium shot,
+bottom-left low angle, bottom-right high angle.
+All expressions different, synchronized movement, vertical format.
+```
+
+> ⚠️ Kling may struggle with consistency. Better to generate separately and combine.
+
+---
+
+#### 13–15 sec: Finale + CTA — Return to Frontal
+
+**Text-to-Video:**
+
+```
+Woman, frontal close-up, delivers punchline directly to camera 
+with confident smile. Slight head nod, wink at the end.
+Static camera, punchy energy, vertical 9:16, TikTok creator vibe.
+Sharp focus, professional ring lighting.
+```
+
+**Image-to-Video:**
+
+```
+Woman smiles confidently, delivers line to camera, 
+subtle wink, slight forward lean. Static shot, no camera movement.
+```
+
+---
+
+### Universal TikTok-Style Modifiers
+
+Add to the end of prompts:
+
+| Goal | Modifier |
+|------|----------|
+| **Vertical** | `vertical 9:16 format, portrait orientation` |
+| **TikTok vibe** | `TikTok creator aesthetic, social media style` |
+| **Lighting** | `ring light, soft natural window light, even lighting` |
+| **Sharpness** | `sharp focus, high clarity, 1080p quality` |
+| **Energy** | `dynamic, punchy, energetic mood` |
+
+---
+
+### Prompt Template for varg SDK
+
+```javascript
+const klingPromptTemplate = {
+  subject: "Young woman, [EXPRESSION]",
+  action: "[MOVEMENT/GESTURE]",
+  camera: {
+    angle: "[frontal/45-degree/low-angle/high-angle]",
+    shot: "[extreme-close-up/close-up/medium-shot]",
+    movement: "[static/slow-push-in/gentle-tilt]"
+  },
+  style: "TikTok creator aesthetic, vertical 9:16",
+  lighting: "ring light, soft natural lighting",
+  negative: "blur, horizontal format, camera shake"
+};
+```
+
+---
+
+## Pro Tips for Scale
+
+- **Consistency via Image-to-Video** — Generate base frame in Flux/SDXL, then animate in Kling. Prevents face "drifting" between angles.
+- **Batch prompts** — Create 3–4 expression variations per angle, then pick the best during editing.
+- **Reference-first** — Kling works better when you provide an image and describe only the movement.