Higgsfield AI × Bloom: The Filmthon

We ran a hackathon where people made videos with AI, and whoever got the most views in one week would win Higgsfield credits. Participants are still posting the videos they made on their own social accounts and working hard to make them go viral.

Minsung Kim (Higgsfield's Korea country manager), Dasol, and I put a lot into preparing this event, but some things on the operations side fell short, and I have some regrets about that. To make up for it, Higgsfield talked with their headquarters and gave all 150 attendees an Ultra plan worth about 200,000 won (around $145) each. It lets you use every feature with no limits, and it adds up to about 30 million won in total. I'm hoping everyone gets an explosion of views over the next week and takes home some of the credit prizes.

The video that Jeongmin from the Bloom community showed on site was really impressive and a lot of fun. We're hoping lots of people take part, and we'll announce the 10 winners once we finish counting.

Fixing the endings you never liked

The filmthon wasn't a sit-and-listen event. Junsoo Kim from our team, who was hosting, opened with a mission. Everyone has a movie ending they didn't like. Use Higgsfield to change it into the ending you wanted. It didn't have to be a movie. A TV drama was fine too.

Just making something and going home isn't much fun, so we made it a competition. Everyone posts the video they made that day on their social accounts, and a week later we rank them by views. Higgsfield put up credits worth 400,000, 200,000, and 100,000 won for first through third place.

About half the room was at Bloom for the first time, so Junsoo gave a short intro to how the community started. Wonjun's post Claude Blue went viral, we reached out to an Anthropic ambassador in Korea to plan the first event, and more than 2,000 people signed up in three days. After seeing that, we decided to run events every week. Last week we did events with Genspark and Datadog.

Maru (a startup hub in Seoul) provided the venue. We always approve everyone for free, so we keep moving between venues to fit in as many people as possible.

When Junsoo asked if there were any designers or video editors in the room, only a few hands went up. Most people didn't work in those fields. Junsoo said it didn't matter: if your views are good, you win.

According to our survey, about 60 percent of attendees had never used Higgsfield.

The rules: pick something you want to remake from an existing movie or drama, or from assets posted in the Higgsfield community. Keep it 30 to 60 seconds. Post it on social media as soon as it's done and tag the Bloom channel.

What if production costs went to zero

Minsung started with the music video for Zephyr, a virtual girl group Higgsfield created. It's a sci-fi K-pop video with five characters, and it took five days to make.

Launching a girl group usually takes a year at minimum and two years at most, and movies are similar. With that comparison set up, Minsung asked the room:

What if all these costs you can't avoid went to almost zero?

A few numbers framed where things stand. 91 percent of companies use video for marketing, 82 percent of all internet traffic is video, and people watch 100 minutes of video a day. But the people making that content hit walls on cost, speed, and scale. Even with a great idea, a movie needs a certain amount of money no matter what, and even if one succeeds, the sequel gets stuck on time all over again.

Minsung also shared Higgsfield's own numbers: 30 million users worldwide and about $700 million in revenue, 17 months after launch. Korea ranks second or third in the world on both revenue and the volume of content made.

It's a platform that gives anyone with an idea, anyone with creativity, the ability to put their work out into the world.

Higgsfield gets called an orchestrator, or an aggregator, because you pick engines like Seedance, Kling, and Nano Banana from a single screen. People often ask which engine is best. Minsung's answer was that it depends on what you're making, and a single 3 to 5 minute piece usually uses four or more.

Game developers came up too. Once you've built a game with vibe coding, the next wall you hit is image assets. The AI wrote the code, but you're stuck because you have no art. Minsung explained that you can make all of those assets inside Higgsfield.

After the talk, Junsoo said that even though we sometimes run three events a week, we hadn't had a crowd this big in a while, which showed how hot Higgsfield is right now. Junsoo also admitted to never having used it yet, and suggested we all practice together with what we'd just learned.

Making the AI do the work

Dasol Kim led the hands-on session. Dasol runs HAXAI, an AI video agency, goes by "Ro Sajang" on YouTube, and used to work as a planner at an ad agency. So Dasol knows firsthand how many steps and how many people it takes to go from planning and copy to production, CG, and VFX.

Asked why Higgsfield, Dasol answered with one comparison.

It's the Netflix of AI. Instead of buying movie tickets one at a time, you subscribe once and get to use all the image and video AI inside.

Dasol calls it the Higgsfield Universe. It's known for shipping new features every week, and Dasol said it's a surprise every time they log in to see how fast things move.

Dasol teaches a lot of workshops, and people tend to forget everything once they're over, so this session was built around leaving people with one keyword: FLOW.

Every good AI video has a flow.

F is Find: find references. L is Logic: plan the story and scenes. O is Open: choose a studio. W is Work, but with the meaning flipped.

We're not going to do the work. We're going to make the AI do the work.

The Find part ended up being the most used during the hands-on time. The Higgsfield community has videos other people posted, and when you open a project, you see every asset used to make it. Characters, outfits, locations, and props are each organized separately, and when you click a scene, you see the prompt, the model used, and even the resolution.

You can steal the know-how of people who are really good at this.

Dasol said this is especially useful if you're struggling with consistency. To put the same character in several scenes, you need shots from many angles, from the front to the side, and looking at the assets others uploaded makes that clear. Dasol explained it follows the same order as casting a film: hire actors, have a stylist pick outfits, settle on locations, prep props. Another trick Dasol uses often is copying a prompt as is and handing it to whatever LLM you use with "change this to fit mine."

After watching one video from the community that day, Dasol said they honestly couldn't tell whether it had been shot in Hollywood. It was a surprise to realize you can now make a whole film with just Higgsfield.

There was a small moment while Dasol was walking through sign-up on screen. Dasol was telling people Higgsfield is in English so they could use Google Translate, and Minsung pointed out that Korean was now supported. The feature had gone live three days earlier. Someone who uses it every day hadn't noticed.

I use Higgsfield every single day, and I didn't know this.

Finding "attractive and likable"

For the Logic step, Dasol used Higgsfield's Supercomputer. It's a chat feature where you pick a model and talk to it. Claude Fable 5.1 had just been added to the list that day, but it cost about twice as many credits as the GPT option, so Dasol recommended the cheaper one to participants who had to make 300 credits last all day.

Dasol's pick was GPT 5.6 Sol. Each use costs about 6 credits, so 300 credits goes a long way.

The prepared demo was a Japanese teen romance short drama. Dasol started with "I want to make a 30-second video, what should I do?" and the model suggested building the world first, then kept offering options for the setting and the male and female leads. Dasol picked a small seaside town, and after just a few clicks, out came character profiles, a storyline, and even narration. No prompt writing needed.

When Dasol asked for character images, the Supercomputer switched on its own to the Soul model, which makes realistic faces that don't look AI-generated. But it did realism so well that the result wasn't what Dasol had in mind for a high school drama.

He was wearing a school uniform, sure, but he came out looking kind of like a middle manager at an office.

Dasol asked again for someone younger since he's a high schooler, and it changed right away. But asking for him to be really good-looking pushed it to an idol look. After a few rounds of adjusting, this was the phrase that worked:

The in-between adjective I found was "attractive and likable."

The model hadn't gotten it wrong. Dasol just hadn't told it which direction to go yet. Give it a clear direction and it delivers, so Dasol said not to panic if the first result isn't what you wanted.

Dasol was candid about prompts too.

I have the LLM write all my image generation prompts. I don't write them myself.

Picking a studio

O is choosing a studio. If you don't even know what to make, scroll down on the Supercomputer screen and you'll see a list of templates. To advertise your company's product, there's Ad Multiplier. For videos that stand out on social media, there's Viral Presets. Animation also starts with a single click.

For people with a product, Dasol recommended the Marketing Studio. You don't need to write prompts at all. You upload your product and Higgsfield composites it in.

If you need product photos, uploading the product is enough to composite it, and you can also train it on the product itself so it shows up consistently across many scenes.

A fair number of people in the survey wanted to make a film, and for that there's Cinema Studio, the most detailed of them all. You upload an image of your lead as a reference, pick a genre like drama, comedy, or horror, choose whether it's modern or a period piece, and even set the video's pacing ahead of time. You can choose the camera too, so if you want a camcorder feel, you pick that, and you set lighting and color palette separately. In the prompt, you use the @ sign to call up a model you made earlier. In the demo, Dasol tagged a character saved under the name "Minam-i" (roughly "Handsome" in Korean) and made a scene of running into the male lead, Haruto.

Dasol had a tip for making video. Don't render in 1080p from the start. Test at 480p, then raise the quality if it looks good. Since length costs both credits and wait time, start short, at 4 seconds. If you turn audio on, it adds background music, and if you put dialogue in double quotes, it even dubs it.

Dasol added that the 4-second advice depends on the model. Seedance 2.5 can produce 30 to 40 seconds in one go. But longer clips cost more credits and take longer, so Dasol suggested testing in 10-second steps.

For people who want to set up images separately, Dasol recommended different models. If people need to look realistic, use Soul. If you need to put your product in and keep it consistent, use Nano Banana. From an image you make, you can click convert to video right away and bring that scene to life.

If writing dialogue yourself feels hard, go back to the Supercomputer and ask it to write all the dialogue and scene descriptions. Dasol shared the slides by QR code and said any comments left on that doc would be checked live from the front during the hands-on time. At the bottom of the doc was also a contact form button for people who want to bring this into their company's work.

We played the rehearsal video

On the way into the hands-on session, the promo code that was supposed to add 300 credits didn't apply right away. While we issued a new code, reposted it on Discord, and waited for confirmation from headquarters, time slipped by.

It was an awkward situation, but Dasol filled it naturally. When the screen wouldn't advance, Dasol went back to the prepared slides. Since showing generation live would mean more waiting, Dasol kept explaining with results made earlier. Dasol had even recorded the rehearsal just in case, and played that. Dasol also shared a doc summarizing the talk, so people could read along while they waited.

The credit limit was originally set for 100 people, but 150 showed up, so Minsung made a request at the start. Anyone who already had credits in their Higgsfield account, please use your own instead of the promo code, and Higgsfield would pay you back with 500 credits within the week. Six of the ten people who already had credits gave theirs up.

You got the promo code by filling out a sign-up form in a temporary channel we made on Discord.

In response, we talked with Higgsfield headquarters and gave all 150 attendees an Ultra plan worth about 200,000 won each, and expanded the prizes from the top 3 to the top 10.

When you run events, mistakes like this happen. We're still learning as we go, running these every week, and next time we'll check the codes first.

A dog that does your work for you

Even so, people made things, and we watched them together in the last 20 minutes.

First we watched Jeongmin's video. Someone asked if it was made that day, Jeongmin said yes, and people wanted to see it again, so we played it a second time.

Next, Ingyu Choi came up front. Ingyu said they hadn't expected to come up and talk, and when asked about the topic, answered:

Work just annoys me every time, so I wished my dog would get things ready for me, and I made that.

The mission was to change an ending, but instead of a movie, it was a video of a dog finding backup files. Ingyu hadn't touched the sound at all, and it even had a save sound effect.

Minsung showed one too, using the Marketing Studio from earlier. With no prompt, just a link to a product page, Minsung took a client's sunglasses, picked an avatar, uploaded two product photos, and ordered a 15-second video.

In the video, the model says the lenses block UV and blue light, that they're so light you forget you're wearing them, and that the gray color is unusual and changes the whole vibe. All of that came from reading the reviews on the product page. The backdrop was Namsan (a mountain in central Seoul). Nobody asked for that. Because the account was using Korean and a Korean IP, it chose that on its own.

It used 4 credits.

The price came down, but the adjectives didn't

Everyone will post their videos on social media, and we'll rank them by views next Thursday. Post them anywhere, YouTube, Instagram, or LinkedIn, and tag the Bloom channel so we can find them.

Thanks to the 150 people who made something in two hours, and to the six people who gave up their credits.

Originally written in Korean on Brunch, September 4, 2026.