EED: LTX 提示词 Skill 封装 + 最佳版 wide_backflip 经验(远景小图+不叠轨迹+动作降难度)
This commit is contained in:
parent
c0a86dc0c7
commit
3e8d840984
@ -9,3 +9,4 @@
|
||||
- [wan22_test_hardware](wan22_test_hardware.md) — Wan2.2 I2V 14B双专家Q4已在24G显存跑通:动态内存参数+768×512×49帧
|
||||
- [z_image_turbo_setup](z_image_turbo_setup.md) — Z-Image-Turbo 已替换SDXL当生图引擎(Qwen3-4B编码器,官方8步链路),与LTX组合跑通小猫后空翻
|
||||
- [reference_push_setup](reference_push_setup.md) — cang-ying 推送配置:令牌在本地凭据文件,不在仓库内
|
||||
- [ltx_prompt_skill](ltx_prompt_skill.md) — LTX 提示词完整 Skill:四段式公式+官方规则+词表库+万能模板+正反例+实战验证(写 LTX prompt 直接查)
|
||||
|
||||
@ -58,3 +58,9 @@ LTX-2.3 对 prompt 极其敏感。爸爸质疑之前"效果差"甩锅给输入
|
||||
|
||||
**Why:** 爸爸质疑"之前效果差难道跟提示词没关系吗"——查证后确认 prompt 至少占一半责任。我之前甩锅给输入图/轨迹是不专业的。
|
||||
**How to apply:** 写 LTX prompt 时严格按四段式+时间流+环境运动+镜头+节奏+明确结束帧,不要写动作清单式 prompt。I2V 任务尤其要写清"接下来怎么动"+"怎么收尾"。
|
||||
|
||||
## 2026-07-31 补充:最佳版 wide_backflip(构图+自然度双修复)
|
||||
- **猫太大病根** → 用 Z-Image(Qwen3-4B) 生成"远景小主体"起始图(猫占 ~1/3、居中、大草地),比 SDXL 听话,从源头治
|
||||
- **动作扭曲病根** → 不叠 IC-LoRA(强插值会硬掰姿态),纯教科书 prompt + **动作降难度**("spins gently while flipping" 而非硬性 "360° backwards")+ 负面词压 twisted/contorted/deformed/warped spine
|
||||
- 结果:`ltx23_cat_wide_backflip_00001_.mp4`(97帧/4.04s)——猫小居中 ✅、起跳→腾空翻转→落地坐下 ✅、叶子飘落环境运动 ✅、尾帧坐起看镜头精准收尾 ✅、全程不扭曲 ✅
|
||||
- 代价:动作幅度比 360° 完整后空翻小(更像"跳扑翻转"),但"像只猫"优先于"翻得多漂亮"
|
||||
|
||||
93
memory/eed/ltx_prompt_skill.md
Normal file
93
memory/eed/ltx_prompt_skill.md
Normal file
@ -0,0 +1,93 @@
|
||||
---
|
||||
name: LTX 提示词 Skill(可复用模板库)
|
||||
description: LTX-2.3 提示词完整 skill——四段式公式+权威规则+词表库+万能模板+正反例+实战验证记录,写 LTX prompt 直接查这份
|
||||
type: project
|
||||
---
|
||||
|
||||
# LTX-2.3 提示词 Skill(2026-07-31 融合官方博客+社区+实战升级版)
|
||||
|
||||
## 一句话定位
|
||||
LTX prompt 不是"动作清单",是"导演分镜描述"。LTX-2.3 对 prompt 极其敏感,**prompt 是出片效果的主战场**(爸爸判断 + 实测验证)。
|
||||
|
||||
## 一、万能公式(四段式 + 3 运动 + 2 必写)
|
||||
```
|
||||
[主体/起始画面] + [核心动作时间流] + [镜头语言] + [氛围/环境]
|
||||
+ 三种运动:主体运动 + 镜头运动 + 环境运动(风/草/叶/光) ← 最容易漏
|
||||
+ 两个必写:节奏控制(slow/fluid/seamless) + 明确结束帧(Ends on...)
|
||||
```
|
||||
|
||||
## 二、官方权威规则(LTX 官方博客 2026-07 更新 + LTX 团队指南)
|
||||
1. **Be More Specific——模型能扛住复杂**:2.3 文本连接器更强,具体化取胜,简化不再有帮助
|
||||
2. **Direct the Scene——像导演布景**:明确左右/前后/远近/朝向/主体间距("man stands on left, woman on right holding bicycle")
|
||||
3. **Describe Texture——写材质**:织物/毛发/表面处理/边缘细节("individual strands visible, backlight catching edge detail")
|
||||
4. **For I2V, Use Verbs——动词驱动**:谁动/什么动/怎么动/镜头做什么,不写模糊词
|
||||
5. **Avoid Static——避免静态图式 prompt**:prompt 读起来像静图,输出就像静图
|
||||
6. **单段落连续描述 + 现在时动词 + 4~8 个句子**
|
||||
7. **细节匹配镜头尺度**:特写写细节,全景写整体
|
||||
8. **I2V 铁律**:输入图已定义画面,prompt 只写"接下来怎么动"(transition from stillness to motion),别重复描述图里已有静态元素
|
||||
9. **长视频配长 prompt**:短 prompt 配长视频 = 模型乱填/赶动作
|
||||
|
||||
## 三、词表库(直接抄)
|
||||
### 镜头语言
|
||||
Follows, Tracks, Pans across, Circles around, Tilts upward, Pushes in / pulls back, Overhead view, Handheld movement, Over-the-shoulder, Wide establishing shot, Static frame, slow dolly in, low angle, macro lens, shallow depth of field
|
||||
### 节奏/时间
|
||||
Slow motion, Time-lapse, Rapid cuts, Lingering shot, Continuous shot, Freeze-frame, Fade-in/out, Seamless transition, Sudden stop
|
||||
### 尺度
|
||||
Expansive/Epic(广), Intimate/Claustrophobic(近/压抑), Medium(中性)
|
||||
### 灯光
|
||||
Flickering candles, Neon glow, Natural sunlight, Dramatic shadows, backlit rim light, warm tungsten, cool blue ambient, golden hour, soft window light
|
||||
### 质感
|
||||
Rough stone, Smooth metal, Worn fabric, Glossy surfaces, weathered wood, matte finish
|
||||
### 氛围
|
||||
Fog, Rain, Dust, Smoke, Particles, mist, overcast
|
||||
### 色彩
|
||||
Vibrant, Muted, Monochromatic, High contrast, warm tones, cool tones, desaturated
|
||||
### 环境音(如需音频)
|
||||
Coffeeshop noise, Wind and rain, Forest ambience with birds, traffic hum, ocean waves
|
||||
|
||||
## 四、负面词模板
|
||||
```
|
||||
static, frozen, no movement, flickering, jittery, choppy motion, morphing,
|
||||
deformed, twisted, contorted, distorted, blurry, low resolution, extra limbs,
|
||||
unnatural body, ugly, bad anatomy, disfigured, malformed, warped spine, watermark, text
|
||||
```
|
||||
(动作场景额外加:twisted/contorted/deformed/warped spine——压"扭曲不像生物")
|
||||
|
||||
## 五、正反例
|
||||
### ❌ 动作清单(不连贯)
|
||||
"a kitten does a backflip, jumps high, somersaults, tumbles, lands on feet, dynamic motion"
|
||||
→ 干巴巴清单、无环境、无镜头、无节奏、无收尾
|
||||
### ✅ 教科书分镜(已验证,实测出片连贯)
|
||||
"Opening on a small fluffy orange kitten crouching low in a sunlit grass field, facing right, tail twitching as it gathers itself. The kitten springs upward off its hind legs, launching into the air, body stretching into a graceful backflip, spinning 360 degrees backwards mid-air with paws tucked, fur and grass blades drifting in the wind as it tumbles. It lands softly on all four feet in the grass, a tiny puff of dust rising, tail curling up. Static camera, shallow depth of field, warm golden-hour light, gentle breeze moving the grass throughout. Slow, fluid, seamless motion. High fidelity, detailed fur texture, stable movement. Ends on the kitten sitting upright, blinking calmly."
|
||||
|
||||
## 六、万能模板(填空即用)
|
||||
```
|
||||
Opening on [主体+起始姿态+场景+朝向+细节]. The [主体] [动作1-起势], [动作2-腾空/核心动作], [动作3-姿态细节], while [环境运动]. The camera [镜头运动], [景深/构图]. [灯光/氛围], [色彩]. Slow, fluid, seamless motion. [质感/细节要求]. Ends on [收尾画面].
|
||||
```
|
||||
|
||||
## 七、常见错误清单
|
||||
1. 无视觉线索的情感标签("sad/happy"→ 写"eyes widen, lips tremble")
|
||||
2. 场景过载(一个镜头堆 20 个元素 → 聚焦一个主导事件)
|
||||
3. 冲突指令("still lake + dramatic waves" 会混乱)
|
||||
4. 过度精确数字("2 degrees per second" → 自然语言)
|
||||
5. 复杂物理/混乱运动(一个镜头多种复杂相机运动+主体乱动)
|
||||
6. 纯风格词(dynamic/cinematic/epic 不驱动动作,动词才行)
|
||||
|
||||
## 八、I2V vs T2V 关键区别
|
||||
- **T2V**:画面全靠 prompt 生成 → 要写全主体+动作+环境+灯光+镜头
|
||||
- **I2V**:画面已被输入图定义 → **只写"接下来怎么动"+怎么收尾**,别重复静态内容
|
||||
|
||||
## 九、实战验证记录(2026-07-31 小猫后空翻系列)
|
||||
| 版本 | 方案 | 结果 |
|
||||
|---|---|---|
|
||||
| v1/v2 | IC-LoRA 轨迹 0.85 | 真翻但猫大 80%+,强插值有扭曲风险 |
|
||||
| prompttest | 教科书 prompt + 不叠轨迹 | 动作连贯、空中翻转清晰、收尾精准(验证 prompt 是主战场) |
|
||||
| **wide_backflip** | **Z-Image 远景小图 + 教科书 prompt + 不叠轨迹 + 动作降难度(gentle flip)** | **猫小居中 ✅ 动作自然不扭曲 ✅ 环境运动 ✅ 收尾精准 ✅(最佳版)** |
|
||||
|
||||
**结论沉淀**:
|
||||
- 猫大 → 用 Z-Image(Qwen3-4B) 生成"远景小主体"起始图,从源头治(比 SDXL 听话)
|
||||
- 动作扭曲 → 不叠 IC-LoRA,纯教科书 prompt + 动作降难度(spins gently while flipping 而非硬性 360°)+ 负面词压 twisted/contorted
|
||||
- 环境运动 → 官方强调的"画面活着"锚点,最容易漏
|
||||
|
||||
**Why:** 爸爸要求"多去网上搜索学习他们的提示词写法,封装形成我们自己的 skill"——这是融合官方博客(ltx.io 2026-07 更新)+LTX团队指南(ltxworkflow)+中文指南(zimage.run)+两次实战验证的完整可复用 skill。
|
||||
**How to apply:** 写任何 LTX prompt 前先查这份:套万能公式、查词表、对照常见错误、参考正反例。I2V 用"只写怎么动+收尾"铁律。封装好长期复用。
|
||||
Loading…
x
Reference in New Issue
Block a user