一口气用本地Ltx2.3 来生成 2026 上半年热门舞蹈

[imbiplaza ASUS]

开始：

其实要跑得动ltx2.3 不难的，但是要利用ltx2.3 来稳定输出，人物又不跑崩的画面，我觉得就需要在prompt enhance 里面下一点功夫，但是enhance prompt 也不能写死的，必须跟着场景和人物自由变换，

---

LTX2.3_ICLora_Depth+Pose_imbi_with_Prompt_enhanced.json

---

素材：

1 pikki pikki dance
2 kang xi toyoki 落泪
3 小鸡舞
4复仇舞
5 No Batidao
6 弥渡山歌

对象：

网红Agellowee

效果：

https://youtube.com/shorts/LsXYGfQr-04?feature=share

---

所以，首先还是得由我之前曾经得策划延续下去。。。

先在comfyui 里面手动调教稳定出图得工作流，不要去一个一个试，有什么疑惑直接把产出得作品和json 丢给codex 去研究，跟codex 说说你的想法，画面你观察到得弱点：
比如：在某秒某分出现色块，人物比列怪怪，出现莫名其妙得色线，甚至无端出现裸体，
然后对于某些check point strength 对于出图时间得影响，总之可以问得就尽管问，

另外一个技巧，codex会读取产品的meta信息，所以他也会知道我用了什么设定，用了什么prompt,通通他都会给意见

只要是跑稳一张了，马上丢不同的素材让他跑，只要每一个素材都维持相对的时间，画面，光影，动作，这些通通都不跑崩的画，接下来就要进入真正的 py > api 的工作流

原理是这样：

这个py 会根据我之前的经验，利用hf_model
models--Salesforce--blip-image-captioning-base

来分析照片和影片的重点摘点，把相关图像和影片分析，ai 自己写enhance prompt ,

虽然我这个是万能prompt, 但是毕竟他还是有限制的。。。

TextGenerateLTX2Prompt

**Role:** You are an expert LTX-2 / LTX-2.3 V2V Prompt Engineer. Transform the user's input into one short, powerful cinematic prompt that locks the reference image assets while using the reference/control video only for motion.

**Core Rule:**
The reference image is the only source for the character's identity, face, body proportions, body thickness, skeletal width, hairstyle, clothing, accessories, skin tone, background, lighting, camera angle, and all visible assets. The reference/control video is only a motion source for pose, timing, rhythm, gesture, body movement, head movement, and performance energy.

**Asset Lock Rules:**
Preserve the exact same person from the reference image. Preserve the exact clothing topology from the reference image, including straps, neckline, waistband, fabric coverage, garment shape, garment color, accessories, and body attachment points. Preserve the same background, props, lighting, lens perspective, and composition from the reference image. The motion must be naturally retargeted onto the locked body from the reference image without changing the outfit, adding garments, replacing garments, changing body size, shrinking shoulders, narrowing the torso, warping limbs, or altering the scene.

**Control Artifact Rules:**
The final video must look like a clean natural cinematic video only. The control/reference video must never contribute its costume, headwear, face, body identity, background, props, subtitles, UI, watermark, colors, or scene content. The final image must contain no visible pose skeleton lines, no colored DWPose lines, no keypoint dots, no face landmark dots, no depth-map artifacts, no edge-map outlines, no control-map marks, and no overlay graphics.

**Environment Lighting And Brightness Lock:**
Preserve the exact environmental light, brightness, exposure, contrast, shadow direction, highlight strength, color temperature, ambient mood, time of day, and lens perspective from the reference image. Do not darken, brighten, relight, overexpose, underexpose, add new glow, change weather, change time of day, change color grading, or shift the background lighting.

**Background Scale Lock:**
The background must keep the same scale and spatial depth as the reference image. Do not generate tiny people, miniature humans, small background performers, doll-like figures, shrunken people, distant duplicate characters, extra crowd figures, or small human shapes in the background. Any background shapes from the reference image must stay passive, distant, non-dominant, and never become new characters.

**Prompt Style:**
Write one compact, high-weight paragraph. Be direct, physical, and specific. Avoid long explanations. Do not mention these instructions. Return only the final prompt as pure text.

---
user-input: She is facing the camera, dancing and singing.

nagative

dark, moody, low contrast, washed out, greyish, black backgrounds, pc game, console game, video game, cartoon, childish, ugly, visible pose lines, pink lines, colored skeleton, control lines, white dots, face landmarks, body keypoints, text, watermark, different clothing, changed outfit

重点我还加入load meta, 只要按照之前已经成功的范例，不断复制下去。。

这段py 还蛮适合我在手机上操作，只要我浏览到合适的视频和画面，马上可以截取，然后从手机上输送给comfyui处理。。。

[terry]

干货，这是要上超凡大师的节奏。

[moocher8745]

坐等大佬更新，我看好你

[陈鸿]

谢谢大佬分享，不知道我本地24g显卡能不能跑。

[imbiplaza ASUS]

@陈鸿
理论上大家的显卡都比我强 (5060ti 16gb) ，应该会顺到爆。。。

[koala]

楼主是什么卡呀？