ANUO

AI多模态方面从业者，因为热爱所以坚持～
目前状态：随心而动，自然发生

[OAK @_OAK200]

Exploring a new storyboard format: Depth Map Storyboards.

Instead of letting Seedance 2.0 inherit the storyboard's visual style, I used a reference image to define the tone and look, while the depth-map storyboard defines the composition and camera framing.

workflow + system prompt below 🧵

[Decart @DecartAI]

Meet Lucy 2.5, our most advanced Live AI model yet.

Lucy edits videos in realtime, now with more capabilities and greater control.

See how it's being used across streaming, e-commerce, advertising, and more 🧵

[ComfyUI @ComfyUI]

Most AI video work is just "generate and hope."

This is different.

Creator seungho__yeo ( IG ) rebuilt a live drum performance from the ground up.

Using the drummer's actual motion data to redesign background mood, lighting, and spatial depth frame by frame.

The pipeline:

→ Motion Tracking
→ Depth Mapping
→ AI Edit ( Seedance 2)
→ AI Relighting
→ Background Reconstruction
→ Cinematic Color Grading All built inside ComfyUI.

This is what intentional AI production looks like. Not more generations, better direction.

This is the level ComfyUI creators are working at.

[Blaine Brown @blizaine]

Maestro v1.3.0 is out: SCAIL-2 character animation, 100% LOCAL, FREE & EASY!

(NEW) "Recast": swap anyone in a video for your own character. Drop a clip, type who to replace ("the woman", "man in red"), drop a character image, hit generate. Masking is fully automatic.

SCAIL-2 Fast: bundled distill runs 6 steps with no CFG, about 13x faster than stock. Looks great! (used in video below).

New gallery trick: scrub any video to the perfect frame & send it straight to your image gen references.

Free, open source, runs on your GPU. One-click install via Pinokio.

BTW, this new light-theme version of "Golden Hour" shipped in v1.2.0 & is pretty chill.

[沐阳 @yyyole]

古早自媒体神器！
现在应该很多人都不知道！！

前几年，这两个软件养活了一大群短视频博主！！
尤其是混剪和鬼畜赛道的。

[数字生命卡兹克 @Khazix0918]

最近几个月，好基友海辛和阿文一直在几个项目里高强度用Seedance做视频，在踩了很多坑以后，他们总结了一套很实用很详细的心得。

我看完了以后，觉得很受用，我自己也学到很多。

在征得他们同意后，也想分享给大家，因为我觉得对每一个想用AI做视频的人，也都非常有用。

先说一个反直觉的结论，大部分你在X和小红书上看到的Seedance提示词，都没什么用。

那些千转万转的帖子，华丽的prompt小作文，他们团队几乎都测了一遍，结果发现能用的没几个，反而浪费了大量时间。

这些提示词唯一验���的事，就是Seedance是真的不挑食，你写什么它都能生成点东西出来，但跟质量没关系。

其实写一段视频prompt比想象中简单得多，只需要填好四个部分：

1. 主体，就是画面里出现的主要人物。建议用图生图先出一张角色参考图，然后在即梦里新建主体，后面直接@就行，复用率极高。
2. 场景，上传一张场景图就够了。场景图里可以带上你的主角，这样能交代角色和环境的比例关系，不容易出现人物大小失调。
3. 音乐，这条非常重要，不要生成任何BGM，只生成对应的音效。BGM会严重干扰后期剪辑和配乐。
4. 镜头，言简意赅地写清楚每个镜头的景别和发生了什么就行，不需要长篇大论。

一个好的Prompt大概是啥样的，我也把海辛的截图放下面了。

除了Prompt怎么写之外，还有一些非常有用的心得：

第一，不需要往Seedance里放分镜故事板。很多人花十几分钟生成一张华丽的故事板，带镜头运动曲线带各种标注和表格，丢进去让Seedance生成，反而错漏百出。

第二，Prompt不是越长越好。很多人用AI写的提示词又臭又长，里面很多都是多余的描述和情绪渲染，上传过主体和场景参考以后就不需要再用文字重复一遍了，没有意义。而且如果一段Prompt连自己都读不完，那说明你对故事要发生什么根本就不关心。

第三，不要给每个镜头加时间戳，比如[0-3秒]这种东西，很多时候都是安慰剂，模型其实不会真的按时间戳来分配时长，对时长影响最大的反而是镜头的数量，所以不如用镜头01这种方式更好。

第四，不要盲目追求清晰度。目前对提示词响应最好的版本是720p，不是1080p，清晰度最后定稿了用Topaz放大到4K就行了。

而最最最重要的事，就是有了AI和Agent之后，我们的效率大幅提升，那省出来的时间应该花在哪呢？

答案只有一个：
构思你的分镜和故事。

绝大多数的技巧，最终都会被模型的进步所吞噬。
但相信我，你的故事。
它不会。

[Julie W. Design @juliewdesign_]

quick world building 🌱✨

midjourney 8.2 preview

[simeon-sanai @Naiknelofar788]

Seedance 2.0 on Higgsfield

Prompt:
Matrix: HARIN vs Smith Clone Army

Ultra-realistic IMAX cinematic rooftop fight in heavy rain, Matrix-inspired grey-green cyberpunk tone, wet reflective concrete, glass skyscrapers, stormy atmosphere, 35mm film look, no background music — only natural rain, impacts, debris, and wind sounds.

HARIN: black leather long coat, black fitted outfit, oval sunglasses.
Agent Smith clones: identical black suits, emotionless, hundreds of copies.

Scene opens with HARIN flying violently across the rooftop and crashing into a concrete wall, creating cracks, dust, and rain shockwaves. She slowly stands, adjusts sunglasses, calm and focused.

Rapid montage of hundreds of Smith clones flooding the rooftop through rain and puddles. Rooftop becomes fully covered by the clone army.

Main fight sequence: HARIN launches into ultra-high-speed combat with superhuman martial arts. Flying kicks, spinning attacks, aerial strikes, shockwave punches, bullet-time rain ripples, clones thrown into AC units and off the rooftop. Black lightning motion trails, cinematic motion blur, dynamic aerial camera movement.

The clone army overwhelms her, piling into a massive black-suited human mountain. Rain pounds the surface.

A glowing white energy begins building from inside the mass. Concrete cracks, rainwater reverses upward, pressure intensifies.

HARIN briefly appears through the glowing gap, sunglasses intact, exhales sharply: "하!!"

Massive spherical energy explosion erupts outward, launching clones across the rooftop. Concrete tears apart, rain forms a temporary vacuum sphere before crashing back down.

Final ultra-wide aerial shot: HARIN standing alone calmly in the destroyed rooftop under heavy rain, coat slowly settling, surrounded by storm and skyscrapers.

Constraints:

- Single rooftop environment only
- No blood or gore
- No costume changes
- Sunglasses never break
- No subtitles or text
- No background music
- Total duration: 15 seconds

[ChatCut @chatcutapp]

Codex is now a full video editor!

Our new plugin opens a full built-in NLE right inside the Codex app, where you and your agent can edit together, effortlessly.

If you're already paying for Codex, you get this one for FREE.

It's available now. Start using it in 1 minute, by just telling your Codex:

"Turn Codex into a Video Editor, read https://t.co/mMA23r1RgY"

[波妞PONYO @ponyodong]

用 ChatCut 剪辑制作_AI 运镜教学视频
一句提示词，Codex 也能直接剪视频了。🎬

这支片用 59 秒讲清 11 种运镜：
摇镜交代空间，跟拍带入行动；俯仰与升降改变力量和尺度；变焦、Dolly Zoom 拉扯注意力与空间；环绕、旋转、手持制造立体、失衡和紧张。

通过 ChatCut 插件，我在 Codex 中完成了本地素材导入、口播策划、字幕生成、MG 封面、章节导航、混音和逐帧检查，所有内容仍可在时间线上修改。

Codex × ChatCut 连接器及操作手册：
https://t.co/KQmXLxzWUc

直接在 Codex 输入：
安装 https://t.co/YdDdDoyFw5

安装后可以：

1️⃣ 导入本地视频、音频和图片
2️⃣ 剪时间线，清理口头语与停顿
3️⃣ 转录并生成字幕
4️⃣ 添加 B-roll、转场、叠加层和 MG 动画
5️⃣ 生成画面、旁白、音乐与音效
6️⃣ 导出并检查最终成片

One prompt, one editable workflow—from raw footage to a polished video.

#Codex #ChatCut #VideoEditing

[dontbesilent @dontbesilent]

不管你想做答疑群，还是想在直播间给别人答疑，你都要意识到

半数以上的人提问并不是为了获取答案，是为不行动寻找一个足够权威的许可

比如会有人问我，他没有营业执照，如何注册企业专业号

实际上他真实的动机是，他不想去干这件事情。只是他想听我这种 “权威人士” 亲口告诉他，没有营业执照他一定干不了，这样他就可以不用去行动了

所以接下来你的选择就是搞清楚，你要赚情绪的钱，还是赚商业的钱

我选择赚商业的钱，所以对方问我没有营业执照怎么办的时候，我的回答是：你可以有

如果你选择赚情绪的钱，那就抚慰一下对方的心灵

但是请注意，你的客单价一定���低，这些人干不成任何事，高单价，必退款

[dontbesilent @dontbesilent]

skills 免费
作者答疑群仅售 ¥199
今天 18 个小时被动收入 4k 🤣

GitHub 才是搞钱第一阵地

[AI様の下僕 @aigeboku]

GPT5.6 Solがリリースってことで、blender操作させてる😎

もう２時間以上は黙々と作り続けてるんだけど、どこまでこだわるんだyo🤟
You、良いね。

[Oogie @oggii_0]

Crafted with GPT image 2 + Seedance 2.0 on @SocialSight

Prompt: Ultra-realistic cinematic WWII battlefield, single continuous tracking shot, 15 seconds, no cuts. Use the attached first frame as strict visual reference for environment, lighting, atmosphere, and composition. Remove all guide overlays: no blue route line, no numbers, no start/end text, no markers.

Camera begins handheld behind a soldier's shoulder as he charges through muddy trenches in driving rain. Smoke, debris, and bullet impacts throw mud into the air. He weaves past destroyed equipment and dives behind Czech hedgehog barricades alongside allied soldiers. Camera rotates around him briefly — revealing the full battlefield scale: WWII fighter planes crossing storm clouds, smoke columns, tanks advancing, thousands of soldiers pushing toward a massive concrete fortress wall. Camera returns behind him as he moves forward.

He sprints toward the armored line. Tanks roll past on both sides, cannons firing with huge muzzle flashes and shockwaves. Enemy fire from the wall sends sparks and concrete fragments flying. He jumps onto a moving tank, runs across its muddy hull, ducks under fire, leaps off toward the fortress base. Camera follows dynamically with low-angle running shots, controlled explosion shake, cinematic motion blur.

At the wall base, camera tilts upward showing the fortress height. He fires a grappling hook, climbs the wall — boots scraping concrete, dust falling from bullet impacts, sparks around him. As he reaches the top, camera rises over his shoulder revealing a massive hidden military complex beyond: endless rows of tanks, armored vehicles, supply trucks, bunkers, searchlights, fires, smoke columns stretching to the horizon. Final frame pulls to a wide aerial shot showing the true scale of war.

Cold gray-blue color grading, volumetric smoke, wet mud, rain, practical explosions, film-grain texture. No modern weapons, no fantasy elements, no text overlays, no cartoon style.

[Runway @runwayml]

Seedream 5.0 Pro is now in Runway. Generate highly detailed images from a prompt or reference, with legible in-image text in up to 14 different languages.

Try it now at the link below.

[leo 🐾 @synthwavedd]

🚨 BREAKING: ByteDance have announced Seedream 5.0 Pro, seemingly the first model from any lab that aims to compete with GPT-Image 2

https://t.co/mRpFfwdUNg

[Drew Brucker ん @dcbruck]

Colorplay with Midjourney v8.2 [preview]

[rob - comfyui @hellorob]

Claude Fable running ComfyUI workflows through the Comfy MCP is a cheat code.

→ Claude pulled the shot from the web, auto-detected the scene cuts + trimmed it
→ ran my saved depthanything v3 + openpose workflow to build the driving reference video
→ scrubbed every frame for the sharpest one, then character swapped me in with gpt-image-2 exactly how I directed it: wearing a tuxedo, liminal space environment, no background characters
→ wrote the seedance 2.0 prompts + ran the gens
→ stitched together the 3-panel comparison you're watching

The preprocessing + prompt writing is usually most of the work on a clip like this. With the MCP I didn't touch any of it. Didn't need to look at a single node lol.

[Higgsfield AI 🧩 @higgsfield_ai]

Seedream 5.0 Pro vs GPT Images 2.0, both fed into Seedance 2.0.

[Xudong Han @Xudong07452910]

Claude Code 有个很值得养成的收尾习惯：

每次关掉一个长会话之前，先让它写「交接文档」。

直接输入：

「这个会话要结束了。请写一份交接文档存到 HANDOFF.md：我们在做什么任务、已经完成了什么、当前卡在哪、下一步计划是什么、有哪些踩过的坑绝对不要再踩。写给一个完全没有上下文的新会话看。」

下次开新会话，第一句话就是「先读 HANDOFF.md」。

这一步特别适合跨天推进的大任务。

因为 AI 的记忆以会话为界，你的项目以周为界，中间的落差，每次都靠你口头补，又累又漏。

聪明但每天失忆的员工，最需要的不是变得更聪明，是一本工作日志。

这��实是给 AI 建「跨会话记忆」的最低成本方案：不装任何插件，一个 md 文件就够。