@PrajwalTomar_: https://x.com/PrajwalTomar_/status/2110709645115596905
Summary
一篇关于用 Claude(Opus 5.5) 搭配 Higgsfield 工作流,15 分钟自制 30 秒电影感 AI 咖啡广告的完整教程,涵盖角色表、场景表等一致性技巧与成本对比,同时推广作者的付费社区。
View Cached Full Text
Cached at: 10/15/26, 02:12 PM
How to Build a Cinematic AI Ad Studio Using Claude (Full Guide)
A client of mine wanted to shoot an ad for their place. They had already hired a proper production studio, and right before the shoot they asked me one question. Is AI good enough to do this yet?
Honestly I didn’t know. So I worked out a workflow for it, and we built their ad with it.
Their ad took a lot of rounds. The one above is a version I made for this article on a made-up cafe, so I can show you every step without showing their brand. A 30 second cafe ad, and I spent about 15 minutes of my own time on it. Now imagine what it looks like if you give it an hour or two.
Here’s what’s inside this article.
-
Why most AI ads look fake
-
The four sheets you build before any video
-
The seven steps, from character sheet to final cut
-
How to direct it when a shot feels off (the part nobody shows)
-
What it cost next to a studio
-
What still isn’t perfect
Quick heads up before you scroll. This exact workflow is already live as a one-click system inside my community, The AI Head Start, which opens next week. If you’d rather have your AI set it up for your brand, everything about it is here.
Why Most AI Ads Look Fake
Most people open a video model, type something like “cinematic coffee ad, warm light, girl drinking latte” and hit generate.
You get something pretty. And then you try to make a second shot, and her face is different, the cup is different, the cafe moved, and the light is coming from the other side. That’s the moment it stops looking like an ad and starts looking like AI.
The model isn’t the problem. A real film crew would never shoot like that either. They lock the actor, the set, the props and the light before anyone presses record.
I’ve been playing around with Higgsfield a lot because I use it for so much of my work. This workflow is what I landed on after watching a ton of videos and running a lot of my own experiments. So what you’re getting here comes out of thousands of Higgsfield credits and a lot of AI model time.
The Setup
The brand is made up. Ember & Oat, a cozy cafe. One customer, Maya, on a rainy morning. The barista only shows up as hands, so there’s just one face to keep consistent.
The ad is one idea. Maya sits by the rainy window, takes a sip, and every sip takes her somewhere else. First a sunny terrace by the sea, then an oat field at sunset, then back to the cafe. It ends on a 4 second shot of the cup with the brand name.
I didn’t click through Higgsfield myself. I gave the whole workflow to Claude (Opus 5.5) with the Higgsfield connector, and told it to stop and show me the results at three points before spending more. It did the directing, Higgsfield did the filming, and my job was to approve.
Step 1, Give It Your References First
This is the one thing I’d change if I ran it again, so it goes first.
Before anything gets generated, upload what you want it to look like. A photo of the cafe vibe you like, a few coffee brands you like the look of, the kind of person you have in mind.
I only added mine halfway through, after the first cup came out plain. If you give them at the start, you skip a whole round of guessing.
Step 2, Build the Sheets Before Any Video
This is the core of the whole workflow. You build a reference image for every person, place and prop, and you approve each one before a single second of video exists.
-
The character sheet. One close-up of her face, then her outfit from the front and the back, cut off at the neck, all on a plain grey background. The grey keeps the model focused on her face and clothes. The neck-down crops mean there’s only one face on the sheet, so the video model can’t mix them up.
-
The cafe. Empty, shot from a three-quarter angle, with the window seat, the table, the olive tree and the counter all in fixed places. A three-quarter angle gives the camera something to move around, and the fixed landmarks let you check the room didn’t change.
-
The other places. The terrace and the oat field, each with the same table and bench shape, so the match cut on every sip works.
-
The cup. Its own sheet with the name spelled exactly, because the cup shows up in almost every shot.
This is what mine looked like.
Claude made eight options for Maya and picked the eighth. It also tried an extra cleanup pass on her, and that pass made it worse. It turned her back view into a second front view. So it threw the fix away, kept the original, and told me why.
The first four sheets cost 6.8 credits.
Step 3, Stop and Approve
This is where I stepped in for the first time.
Everything looked good except the cup. The branding felt basic. So I sent one message saying that, with a cafe photo I liked and a few coffee brands as inspiration.
It came back with a new brand look, a chunky rounded wordmark in espresso brown, cream and burnt orange, a new cup with the name stacked on it, and a restyled cafe. It also told me it used my photos only as a style guide, so none of those brands’ actual logos got copied.
That whole round, sheets plus the redo, came to about 23 credits.
Step 4, Draft Cheap at 480p
Video is where the credits go, so you never make the final version first. I started with the first 16 seconds, the cafe to the terrace plus the cup shot.
It made the draft at low quality, and before I even watched it, it gave me a list of what to check. There was a faint ghostly figure in the rain window behind her from about 4 to 8 seconds. And the olive tree was on her right when the sheet had it on the left.
That’s the whole point of a cheap draft. It caught two mistakes for 48 credits instead of finding them in the expensive version.
Step 5, Direct It Like a Film
Then I watched the draft, and two things felt wrong to me.
At the start Maya just sits there frozen, like nothing’s happening. And the cut to the cup was so harsh it happens out of nowhere. After that I liked it honestly.
So I told it exactly that, in plain words. And this is the part I didn’t expect. It worked out why on its own.
The first shot had been written as a locked camera with no action, so the video model held her still. The cut had nothing moving to carry your eye into it. It had kept the camera moving across the terrace cut, just not in the opening.
So it rewrote the first eight seconds.
-
The camera never stops. It slowly pushes in past blurred olive leaves.
-
Maya does small things. She breathes, blinks, tucks her hair behind her ear.
-
At 2.4 seconds the espresso machine hisses and she turns toward the counter.
-
The cut follows her eyes to the cup sliding in, so the sound and the look carry you across.
-
It added “no people or reflections in the window” for the ghost, and moved the tree back.
The new draft cost 36 credits and fixed both. The opening breathes now, and the cut feels like a real edit.
If you take one thing from this article, take this. When a shot feels dead, don’t ask for “more cinematic”. Give every shot a small action, keep the camera moving, and give every cut a reason, a look or a sound.
If seven steps sounds like a lot, this is exactly why I built the community. You paste one link into your AI and it runs these steps with you, for your brand. More here.
Step 6, Make the 1080p Version
Once the draft was right, I said yes, and it made the 1080p versions of the scene and the cup shot and put them back to back.
Step 7, Stretch It to 30 Seconds
16 seconds felt short, so I asked for 30. This is where I learned the most.
The first try was one new 14 second clip, and the framing broke at the oat field cut. The second try was three short clips. It worked, but you could see a seam at 12 seconds.
What actually worked was extending the take I had already approved. It continued the same clip forward by 14 seconds, and it got a matching still of Maya sipping in each new place, the terrace, the oat field and the cafe, as a reference. No seam, and she stays the same person the whole way.
One small trick. For the sound it only asked for ambience, so the model doesn’t add sighs her mouth isn’t actually doing.
Then it put the 12 second scene, the 14 second extension and the cup shot together, 29.8 seconds, and made a 9:16 reel for Instagram too.
What It Cost
The first 16 seconds, every option, the rejected cleanup pass, the cup redo, both drafts and the finals, came to 299.28 credits. Stretching it to 30 seconds, with the two failed tries, took another 311.98. So the whole 30 second ad came to about 611 credits.
A studio is a different world. One US studio’s 2026 breakdown puts the average 30 second local business commercial at about $8,000, with planning, a half-day shoot with a three-person crew and editing (C-I Studios).
What Still Isn’t Perfect
-
At 5 seconds she already has the cup raised, so you never actually see her pick it up. A real director would catch that.
-
The first draft was still and abrupt, and it took one retry to fix. Budget for at least one.
-
Going from 16 to 30 seconds took three tries. If you know you want 30, plan for the extension from the start.
-
The cleanup pass on the character made her worse. Not every fix is a fix, so always compare before you keep it.
-
It’s one person and one product. More characters means more sheets and more retries, so budget more credits.
My Honest Take
AI ads have become very realistic, but obviously they can’t beat a production studio. That’s what I told my client. What we could do is work on a workflow together, and this is the workflow I came up with.
For their ad we took a lot more iterations and spent a lot more time. The one at the top took about 15 minutes of my time. The workflow is exactly the same, so if you want something you’d actually hand to a client, plan for multiple rounds.
The Whole Workflow in One Line
References → sheets → approve → 480p draft → direct the fixes → 1080p → extend to 30 seconds
Run it once with a made-up brand like I did, and the second time it’s just swapping the person, the place and the product.
How This Runs Inside My Community Now
Honestly this is the part I’m most excited about.
This exact system is already live inside The AI Head Start, my community that opens next week.
You get one page to read and one link for your AI. Paste the link into Claude, say “set this up for my business”, and it runs the whole thing with you. Your brand, your product, your ad. Still kinda crazy to me that it’s that simple.
That’s how every workflow I test works in there now. You understand the idea, your AI does the setup.
The first 100 members lock in $25 a month for as long as they stay. Everything about it is here https://theaiheadstart.com/x
Similar Articles
@PrajwalTomar_: No way Claude built this entire cafe commercial I connected Opus 5.5 to an AI video studio, handed it my ad workflow an…
A creator demonstrates connecting Claude Opus 5.5 to an AI video studio to autonomously plan and produce an entire cafe commercial from a brief in about 15 minutes, including self-review and revision of the first draft.
@DeRonin_: Made this video with just 1 prompt Higgsfield Supercomputer. 5 minutes. cinematic ad out here's the framework so you ca…
一份使用 Higgsfield Supercomputer 制作电影级 AI 视频的框架指南,强调以电影导演思维进行提示、使用图像到视频技术、风格参考、长时间镜头、负面提示和外部音效设计。
@PrajwalTomar_: https://x.com/PrajwalTomar_/status/2100926059218870580
This article explains a workflow using GPT-6 Astra as an AI director and Higgsfield as a rendering studio to create motion graphics, detailing steps, costs, and comparisons to traditional agencies.
@eternityspring: Since so many people are watching, and some friends in the comments mentioned qwen3.8 27b I'll share the complete workf…
该文章分享了使用qwen3.8 27b、shuohao-skills、Krea-2和MiniMax H3等AI工具进行AI短剧制作的完整工作流程,强调低成本和可自定义性。
@Khazix0918: https://x.com/Khazix0918/status/2110587546271191166
The author shared and deeply explored the open-source Claude Skill "Anatomy," which can generate interactive 3D mechanical principle models from a single sentence, and even further produce popular science videos with explanatory voiceover, while also discussing the transformation of future learning methods in light of Karpathy's views on AI content understanding.