Add creative/comfyui v5.1.0 (official bundled) and media/ltx-video-pipeline v2.0.0 (custom LTX pipeline)
This commit is contained in:
Hermes Agent
@@ -0,0 +1,134 @@
|
||||
# LTX Video Prompting — Quick Reference
|
||||
|
||||
Condensed from the 544-line deep research at `/home/n8n/workspace/general/LTX_prompt_instructions.md`.
|
||||
|
||||
## Prompt Order (Official LTX 2.3)
|
||||
|
||||
LTX Studio's pipelines weight tokens by position. Two orderings work well:
|
||||
|
||||
**Order one (shot-size-first):** framing → subject → action → style → motion
|
||||
Use when the shot size is the most important decision. "Wide shot, low angle, a detective in a long coat walks toward a warehouse door, harsh sodium lighting, slow push-in."
|
||||
|
||||
**Order two (subject-first):** subject → action → framing → style → motion
|
||||
Use when the character or object is more critical. "A woman in a red coat walks through a Tokyo street at night, medium tracking shot, neon reflections in puddles, handheld follow."
|
||||
|
||||
**Failure mode:** piling every adjective into one long noun phrase. Cut to the 2-3 that matter most.
|
||||
|
||||
**For I2V specifically:** Focus the prompt on motion and action — the visual starting point is already defined by the input image. Avoid describing static elements already visible in the image. Describe the transition from stillness to motion.
|
||||
|
||||
Source: ltx.io/blog/ltx-2-3-prompt-guide, ltx.io/blog/how-to-write-a-prompt (official, June 2026)
|
||||
|
||||
## Shot Sizes
|
||||
|
||||
| Shot | What It Shows |
|
||||
|------|---------------|
|
||||
| Extreme Wide | Vast environment, subject tiny |
|
||||
| Wide / Establishing | Full body + surroundings |
|
||||
| Medium Wide | Knees up |
|
||||
| Medium Shot | Waist up |
|
||||
| Medium Close-up | Chest up |
|
||||
| Close-up | Face and neck |
|
||||
| Extreme Close-up | Single feature |
|
||||
|
||||
## Camera Angles
|
||||
|
||||
| Angle | Effect |
|
||||
|-------|--------|
|
||||
| Eye Level | Neutral, familiar |
|
||||
| Low Angle | Power, dominance |
|
||||
| High Angle | Vulnerability |
|
||||
| Bird's Eye | Detached, god-like |
|
||||
| Dutch Angle | Unease, tension |
|
||||
| Over-the-Shoulder | Dialogue |
|
||||
| POV | First-person immersion |
|
||||
|
||||
## Camera Movements
|
||||
|
||||
| Movement | Prompt Phrase |
|
||||
|----------|---------------|
|
||||
| Static | "Static camera," "tripod-locked" |
|
||||
| Dolly-in | "Slow dolly-in," "camera pushes forward" |
|
||||
| Dolly-out | "Camera pulls back slowly" |
|
||||
| Pan | "Slow pan left/right" |
|
||||
| Tilt | "Tilts up to reveal" |
|
||||
| Tracking | "Tracks alongside," "follows behind" |
|
||||
| Orbit | "Camera orbits slowly around" |
|
||||
| Handheld | "Subtle handheld, micro-shakes only" |
|
||||
|
||||
## Lighting
|
||||
|
||||
| Type | Prompt Phrase |
|
||||
|------|---------------|
|
||||
| Window light | "Soft north-window light, gentle falloff" |
|
||||
| Golden hour | "Golden hour rim light, long shadows" |
|
||||
| Overcast | "Overcast, diffuse light, low contrast" |
|
||||
| Tungsten | "Practical lamp as key, warm tungsten" |
|
||||
| Neon | "Neon glow, cyan and magenta reflections" |
|
||||
| Backlight | "Backlit rim light, subject silhouetted" |
|
||||
| High contrast | "High-contrast studio lighting, deep shadows" |
|
||||
|
||||
## Lens Cues
|
||||
|
||||
- "35mm lens, f/2.8, shallow depth of field" — people/products
|
||||
- "85mm portrait lens, bokeh, crisp eyes" — talking heads
|
||||
- "16mm wide, deep focus" — rooms/landscapes
|
||||
- "50mm standard" — natural perspective
|
||||
|
||||
## Motion Verbs
|
||||
|
||||
Walking, pouring, turning, lifting, revealing, drifting, sprinting, reaching, pausing, stepping, rising, falling, swaying, flickering, rotating, gliding, accelerating, decelerating, spinning, floating, sliding, bouncing, nodding, shaking, gesturing, breathing, blinking, smiling, frowning, gasping, laughing, crying, whispering, shouting
|
||||
|
||||
## Emotion — Show, Don't Tell
|
||||
|
||||
❌ "A sad woman sits at a table"
|
||||
✅ "A woman sits at a table, her shoulders slumped forward, eyes downcast, fingers tracing the rim of an empty coffee cup"
|
||||
|
||||
## Negative Prompt
|
||||
|
||||
```
|
||||
no extra limbs, no face warp, no object duplication
|
||||
no text artifacts, no floating logos, no watermark
|
||||
no extreme motion blur, no rolling shutter wobble
|
||||
no flicker, no frame-to-frame texture shift
|
||||
no Dutch angle, no rapid handheld, keep horizon level
|
||||
```
|
||||
|
||||
For I2V: add `static, frozen, no motion, Ken Burns zoom`
|
||||
|
||||
## Transition LoRA (zhuanchang)
|
||||
|
||||
- **Use ONLY on transition/morph scenes** — character morph, style change, scene switch, day→night
|
||||
- **NEVER on standard I2V** — the LoRA forces transformation behavior even without the trigger word. Confirmed 2026-07-21: caused subject duplication and wrong-scene hallucination on standard I2V scenes.
|
||||
- **Place at END of prompt**
|
||||
- **CFG: 4.0** (not 1.0 — Transition LoRA needs higher CFG)
|
||||
- **Strength: 1.0**
|
||||
- **Node: LoraLoaderModelOnly** (standard ComfyUI node, NOT LTX2LoraLoaderAdvanced)
|
||||
|
||||
## guide_strength (RESOLVED 2026-07-22)
|
||||
|
||||
**Use 1.0 for all standard I2V with ID LoRA dropped.** The old "1.0 = NEVER" rule was from v4 where ID LoRA was also at 1.0 — the duplication was stacked LoRA interference, not guide_strength alone. With ID LoRA dropped, 1.0 is clean and provides proper start-frame anchoring.
|
||||
|
||||
| Value | When | Why |
|
||||
|-------|------|-----|
|
||||
| **1.0** | Standard I2V (no ID LoRA) | Proper start-frame anchoring. Confirmed clean in 2-clip test. |
|
||||
| **0.5-0.7** | I2V WITH ID LoRA (talking-head only) | Lower strength prevents LoRA interference |
|
||||
|
||||
Verify after render:
|
||||
```bash
|
||||
ffprobe -v quiet -show_entries format_tags=prompt output.mp4 | grep -oP 'guide_strength.*?(\d+\.?\d*)'
|
||||
```
|
||||
|
||||
## Common Mistakes
|
||||
|
||||
| Mistake | Symptom | Fix |
|
||||
|---------|---------|-----|
|
||||
| guide_strength too low (0.5) | Identity drift, weak start-frame anchoring | 1.0 (safe with ID LoRA dropped) |
|
||||
| ID LoRA on action scenes | Frontal-face bias, fights camera motion | Drop ID LoRA entirely |
|
||||
| Transition LoRA on standard I2V | Wrong scene, hallucinated content | Remove Transition LoRA entirely |
|
||||
| zhuanchang on every scene | Model forces morphing everywhere | Only on actual transitions |
|
||||
| No camera direction | Static/random movement | Add dolly/pan/tracking |
|
||||
| Prompts too short (<30 words) | Model fills gaps with random data | 50-80 words minimum |
|
||||
| No motion detail | Character freezes | Sequential action in present tense |
|
||||
| Emotional labels | Abstract expressions | Physical cues only |
|
||||
| Conflicting descriptions | Averaged competing signals | One speed, one camera, one light |
|
||||
| euler_ancestral_cfg_pp on fp8 distilled | Catastrophic artifacts, hallucinated subjects | Plain euler only |
|
||||
Reference in New Issue
Block a user