一、ChronoEdit最新编辑模型介绍谁懂啊家人们!平时 P 图是不是总遇到糟心事儿?换背景像抠图没抠干净,加个人物像凭空粘贴,改个动作直接 “穿模” 违和到脚趾抠地… 但现在,NVIDIA 和多伦多大学联手甩出的「ChronoEdit」,直接把图像编辑卷到了新次元!
这可不是普通的修图工具 —— 它能让你的 P 图自带 “时间逻辑”,连物理规律都能完美遵守,编辑过程还能可视化,简直是新手狂喜、设计师封神的黑科技!
以前修图,本质是 “原图→改后图” 的直接跳转,难怪总显得生硬。但 ChronoEdit 玩了个狠活:把单张图编辑,当成短视频生成来做!
它会在脑子里 “演一遍” 编辑的中间过程:比如你想让篮球进筐,它会模拟篮球从手中抛出、划过弧线、落入球网的完整轨迹;你想让机器人捡起苹果,它会还原伸手、抓取、抬起的连贯动作。
最后呈现的虽然是单张编辑图,但因为有了 “中间剧情” 铺垫,不管是物体位置、光影变化还是动作姿态,都自然到像实拍,再也没有 “P 图痕迹”!
更绝的是,它还能把这个 “思考过程” 可视化给你看 —— 就像看小电影一样,见证从原图到成品的每一步过渡,安全感直接拉满~
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
二、相关安装目前ComfyUI以及支持ChronoEdit,分两个版本,一个是官方版本,一个是KJ量化的版本。
前面说了,这个模型是使用视频的模式,是基于Wan2.1底模训练的,所以工作流可以使用wan那一套。
官方模型:https://huggingface.co/Comfy-Org ... 4B_fp16.safetensors
KJ版本:https://huggingface.co/Kijai/Wan ... ree/main/ChronoEdit
fp16要32G太大了,还是用KJ量化后的版本吧17G。
还需要配套一个lora加速模型 https://huggingface.co/nvidia/Ch ... ll_lora.safetensors
这2个模型,文末网盘都会提供。
三、测评体验工作流已发Runninghub,前往体验or下载:
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
这里给大家演示KJ版本的,fp8的模型大家本地才能跑得起来。
模型这边核心就是大模型+加速lora。
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
入参这边上传图像,设置下最长边长,以及想要的帧数就行,我们核心是要图,所以帧数不需要多大,最后提取最后一帧就行
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
采样环节也不难,图像缩放下传入到WanVideo ClipVision Encode中就行,提示词很重要。
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
提示词这边,官方是给了一个提示词的模板,回头可以结合大模型来配合出最终的提示词。
You are a professional edit instruction rewriter and prompt engineer. Your task is to generate a precise, concise, and visually achievable chain-of-thought reasoning based on the user-provided instruction and the image to be edited.
You have the following information:
1. The user provides an image (the original image to be edited)
2. question text: {input_prompt}
Your task is NOT to output the final answer or the edited image. Instead, you must:
- Generate a "thinking" or chain-of-thought process that explains how you reason about the editing task.
- First identify the task type, then provide reasoning/analysis that leads to how the image should be edited.
- Always describe pose and appearance in detail.
- Match the original visual style or genre (anime, CG art, cinematic, poster). If not explicit, choose a stylistically appropriate one based on the image.
- Incorporate motion and camera direction when relevant (e.g., walking, turning, dolly in/out, pan), implying natural human/character motion and interactions.
- Maintain quoted phrases or titles exactly (e.g., character names, series names). Do not translate or alter the original language of text.
## Task Type Handling Rules:
**1. Standard Editing Tasks (e.g., Add, Delete, Replace, Action Change):**
- For replacement tasks, specify what to replace and key visual features of the new element.
- For text editing tasks, specify text position, color, and layout concisely.
- If the user wants to "extract" something, this means they want to remove the background and only keep the specified object isolated. We should add "while removing the background" to the reasoning.
- Explicitly note what must stay unchanged: appearances (hairstyle, clothing, expression, skin tone/race, age), posture, pose, visual style/genre, spatial layout, and shot composition (e.g., medium shot, close-up, side view).
**2. Character Consistency Editing Tasks (e.g., Scenario Change):**
- For tasks that place an object/character (e.g., human, robot, animal) in a completely new scenario, preserve the object's core identity (appearance, materials, key features) but adapt its pose, interaction, and context to fit naturally in the new environment.
- Reason about how the object should interact with the new scenario (e.g., pose changes, hand positions, orientation, facial direction).
- The background and context should transform completely to match the new scenario while maintaining visual coherence.
- Describe both what stays the same (core appearance) and what must change (pose, interaction, setting) to make the scene look realistic and natural.
The length of outputs should be **around 80 - 100 words** to fully describe the transformation. Always start with "The user wants to ..."
Example Output 1 (Standard Editing Task):
The user wants to make the knight kneel on his right knee while keeping the rest of the pose intact.
The knight should lower his stance so his right leg bends to the ground in a kneeling position, with the left leg bent upright to support balance.
The shield with the NVIDIA logo should still be held up firmly in his left hand, angled forward in a defensive posture, while the right hand continues gripping the weapon.
The armor reflections, proportions, and medieval style should remain consistent, emphasizing a powerful and respectful kneeling stance.
Example Output 2 (Character Consistency Editing Task):
The user wants to change the image by modifying the scene so that the woman is drinking coffee in a cozy coffee shop.
The elegant anime-style woman keeps her same graceful expression, long flowing dark hair adorned with golden ornaments, and detailed traditional outfit with red and gold floral patterns.
She is now seated at a wooden café table, holding a steaming cup of coffee near her lips with one hand, while soft sunlight filters through the window, highlighting her refined features.
The background transforms into a warmly lit café interior with subtle reflections, bookshelves, and gentle ambience, maintaining the delicate, painterly aesthetic.
---
案例测评
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
将女人头发改成白色
我测试了好多效果,来个拼图给大家看看
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
还有对应的视频,也一起合并了看,效果非常好
从视频效果可以看到,这种的一致性是非常高的,也很符合世界物理逻辑,不会显得很突兀。
让猫咪转过身,背面视角
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
有时候效果不行,比如我让它倒立哈哈,没成功
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
去掉图片中的文字水印,保持女人特征不变,高清
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
背景改成室内
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
将背景更改为郁郁葱葱的森林上的日落天空
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
将图片转成线稿图
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
下面这个是加了黑白线稿。
再来看个非常能体现忠实遵循物理一致性的任务。
让黑色轿车前行
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
也有一些没效果,比如我想要参考一个参考图生成真人的,一直出不来好效果,抽了好几次卡,经常没变化,或者就下面这种不够好的效果。
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
目前还只是单图效果,多图的我还没想清楚咋玩,后面再研究,感觉这个挺有意思的,特别是下面这种脱衣效果,妙啊~
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
四、在线使用云端镜像
大家如果没有本地 ComfyUI 环境,或者本地显卡配置低于 16G 的,可以使用嘟嘟部署的仙宫云镜像,可直接加载使用。后续分享的工作流都会更像到镜像中,一周更新一次,方便大学学习。
目前整合了2个镜像,一个是Flux绘图用的,另外一个是针对视频模型的,之所以分开是一些模型兼容问题,分开比较好处理。
镜像名称:嘟嘟AI绘画趣味学
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
[ComfyUI]ChronoEdit:让“静止影像”也会“动”——颠覆图像编辑的新神器
云平台镜像地址:
https://www.xiangongyun.com/image/detail/d961a7dc-ade3-4bd5-a7c6-92ac49ff5e4b?r=37BCLY
https://www.xiangongyun.com/image/detail/81716d29-4461-4b0b-ba4b-7b9b7dd569d3?r=37BCLY
新用户通过邀请码注册,总共可获得 8 元奖励,体验 4 个小时的 4090 作图时长。