Updated: Mar 29, 2026
base modelDownload
1 variant available
Click here to try workflows online first:
---
Workflow: LTX2.3 – First & Last Frame Dual Image to Video – Fully Automatic Prompt – Freely Control Last Frame Position – Ultra-High Success Rate
Experience link: https://www.runninghub.ai/post/2035325472770367490/?inviteCode=rh-v1401
Workflow: LTX2.3 – Fully Automatic Prompt – Image to Video – Modular Optimized Version
Experience link: https://www.runninghub.ai/post/2031218459471777794/?inviteCode=rh-v1401
Workflow: AA – Zimage Large Collection – Text to Image + Image Refinement + Inpainting + Outpainting + SeedVR2 Upscaling
Experience link: https://www.runninghub.ai/post/2007084925857570818/?inviteCode=rh-v1401
Workflow: AA – Four-Grid – Storyboard Design – Scene Design
Experience link: https://www.runninghub.ai/post/2026958885080272897/?inviteCode=rh-v1401
Workflow: All-in-One Image Pro – Text to Image – Single, Double, Triple, Quadruple Images – Image Editing
Experience link: https://www.runninghub.ai/post/2026244873988345857/?inviteCode=rh-v1401
Workflow: AA – Various Small Tools for Image, Audio, Video Processing (Continuously Updated)
Experience link: https://www.runninghub.ai/post/2027021102093967362/?inviteCode=rh-v1401
Workflow: AA – FlashVSR Video Upscaling + RIFE Frame Interpolation
Experience link: https://www.runninghub.ai/post/2008534784552738818/?inviteCode=rh-v1401
EN
> A dual‑image (first+last frame) and image‑to‑video workflow designed specifically for LTX2.3. Supports multi‑frame reference (up to 5 images) to generate coherent motion, fixed characters, and consistent scenes for high‑quality AI short films.
Key features
- First+Last Frame Control – Upload a first and last frame, use the index parameter to precisely set the last frame position (intermediate frame strength). This dramatically improves the model’s “directing ability”.
- Multi‑frame Reference – Up to 5 reference images – ideal for e‑commerce makeup changes or complex action sequences.
- Automated Prompt Expansion – Built‑in expand node. However, it’s recommended to manually write clear prompts following “chronological order of actions + scenes + camera angles” so the expansion doesn’t alter your intent.
- Double Sampling + Upscaling – Latent size halved before first sampling, then doubled back before second sampling (default). Ensures high quality.
Recommended settings
| Parameter | Suggested value |
|-----------|----------------|
| Image size | Determined by first frame; scale longest side to 1280 (or 1920) |
| Duration | Target 10s → generate 15s, then trim unusable parts |
| Frame rate | 25fps (cinematic) / 30fps (short videos) / 60fps (high motion) |
| Last frame index | Set to 1 to make the last frame the very end |
| Upscale factor | 3x (1280 → 3840, 4K) |
Compatible tools / nodes
- Base image generation – Zimage (realistic portraits) or “Big Banana” (4K storyboard)
- Storyboard – 4‑panel storyboard workflow → split into single images → optionally upscale to 4K
- Running environment – Runninghub (cloud, multi‑concurrent) or locally (replace prompt expansion node with Qwen3 reverse‑prompt node)
Sample output
- Western Cthulhu short film (80s/90s film texture)
- 11 first+last frame shots + several standard image‑to‑video shots
- ~50% success rate (very high for an open‑source model)
Important notes
- First and last frame images must have identical dimensions
- For complex motions, break them into multiple reference images + simple prompts
- High‑motion, grand scenes, distant characters may still require closed‑source models, but LTX2.3 delivers much more pleasant image quality without the obvious “AI look”
Download workflow (Runninghub JSON / ComfyUI format) ��
[attach file or link]
Try it out and share your results!
---
CN
> 一个专为 LTX2.3 设计的双图生视频（首尾帧）及图生视频工作流，支持多帧参考（最多 5 张图），可生成连贯动作、固定角色与场景的高质量 AI 短片。
工作流特点
- 首尾帧控制 – 上传首帧和尾帧图片，通过 index 参数精确控制尾帧位置（中间帧幅度），让模型“导演力”明显提升。
- 多图参考 – 支持插入最多 5 张参考图，适合电商换妆、复杂动作序列。
- 提示词自动扩写 – 内置扩写节点，但建议按“时间顺序描述动作 + 场景 + 镜头角度”手动写清楚，扩写不会改变原意。
- 二次采样 + 放大 – 一次采样前尺寸缩放 1/2，二次采样后放大 2 倍（默认），保证画质。
推荐参数
| 参数 | 建议值 |
|------|--------|
| 图片尺寸 | 首帧决定比例，按最长边缩放到 1280（或 1920） |
| 时长 | 目标 10 秒 → 生成 15 秒，剪掉不可用部分 |
| 帧率 | 25fps（电影感）/ 30fps（短视频）/ 60fps（高动态） |
| 尾帧位置（index） | 设为 1 时尾帧即为最后一帧 |
| 放大倍数 | 3 倍（1280 → 3840，4K） |
配合使用的节点 / 工具
- 底图生成 – Zimage（真人）或“大香蕉”全能图片（4K 分镜）
- 分镜设计 – 四宫格分镜工作流 → 拆分单张 → 可选放大到 4K
- 运行环境 – Runninghub（云平台，支持多并发）或本地（需将提示词扩写节点替换为千问 3 反推节点）
成品参考
- 西部克苏鲁风格短片（80/90 年代电影质感）
- 11 个首尾帧镜头 + 若干图生视频镜头
- 抽卡成品率约 50%（开源模型中很高）
注意事项
- 两张首尾帧图片**尺寸必须一致**
- 复杂动作尽量拆解为多张参考图 + 简单提示词
- 高动态、大场面、人物远景建议仍用闭源模型，但 LTX2.3 画质更舒服、不显“AI 味”
下载工作流（Runninghub JSON / ComfyUI 格式）��
[附文件或网盘链接]
欢迎试用并分享你的成果！