Fix: flatten comfyui and ltx-video-pipeline to Gitea convention (flat skill dirs). comfyui v5.1.0, ltx-video-pipeline v2.0.0.

2026-07-22 16:19:32 -05:00
parent d94132f4ca
commit d824f865c1
61 changed files with 184 additions and 674 deletions
@@ -0,0 +1,43 @@
+# LTX Prompt Validation Checklist
+
+Run this 10-point check on EVERY segment prompt BEFORE submitting to the render queue. This is a quality gate — do not skip.
+
+## The 10 Checks
+
+| # | Check | Pass if |
+|---|-------|---------|
+| 1 | Shot type | Contains one of: Medium shot, Wide shot, Close-up, Low angle, Tracking shot, Overhead, POV |
+| 2 | Camera direction | Contains one of: pan left/right, dolly in/out, tilt up/down, zoom, tracking, static, handheld, crane, Steadicam |
+| 3 | Actions (2-3 max) | 2-3 distinct actions, present tense, sequential. Not 5-8 simultaneous actions. |
+| 4 | Lighting | Describes light source, quality, and color temp (e.g., "warm afternoon light from tall windows") |
+| 5 | Environment | Describes textures, atmosphere, setting details |
+| 6 | Word count | 50-80 words for I2V (action-focused — start frame already defines visuals). Official LTX says longer is better for T2V, but I2V needs less scene-setting. Under 50 = model fills gaps. Over 80 = model drops actions. |
+| 7 | No text/logos | No words like "text", "logo", "sign", "label", "readable letters" in the prompt |
+| 8 | No zhuanchang | Only append "zhuanchang" if this is a transition/morph scene. Never on standard I2V. |
+| 9 | No emotional labels | No abstract emotions ("sad", "angry", "happy"). Use physical cues instead ("shoulders slumped", "eyes downcast"). |
+| 10 | Single logic | One speed, one camera direction, one lighting logic. No conflicting descriptions. |
+
+## Prompt Structure (Official LTX 2.3)
+
+Two valid orderings per ltx.io (June 2026):
+
+**Shot-size-first:** framing → subject → action → style → motion
+**Subject-first:** subject → action → framing → style → motion
+
+For I2V: focus on motion and action — the start frame already defines the visual. Describe the transition from stillness to motion, not static elements already in the image.
+
+## Example: FAIL
+
+> The camera continues its slow steady pan right through the same neon-lit cyberpunk room, past flickering holographic displays and tangled cables, gradually revealing a woman in her late 20s with chrome temple implants standing by a rain-streaked window. Pink and blue neon light traces her silhouette. She turns her head toward the camera. Dark atmosphere, photorealistic, cinematic 35mm.
+
+**Fails:** Check #1 (no shot type — missing "Medium shot" or equivalent).
+
+## Example: PASS
+
+> Medium shot, slow steady pan right. The camera continues through the same neon-lit cyberpunk room, past flickering holographic displays and tangled cables, gradually revealing a woman in her late 20s with chrome temple implants standing by a rain-streaked window. Pink and blue neon light traces her silhouette. She turns her head toward the camera. Dark atmosphere, photorealistic, cinematic 35mm.
+
+**Passes:** All 10 checks. Shot type present, camera explicit, 2 actions (standing, turns head), lighting described, environment detailed, ~60 words, no text/logos, no zhuanchang, no emotional labels, single direction/speed.
+
+## Audit History
+
+- 2026-07-22: Cyberpunk Clip 2 prompt caught missing shot type in pre-render audit. Fixed before dispatch.