I also made a music video for CAMIEL, keeping the same 1970s look and over-the-top action vibe.

It was fun taking these characters out of the usual dialogue scenes and letting the music carry things for a change.

You don’t need to have watched the episode to check this out. I’m actually curious how it comes across without that context. Does it make you want to see more of the story?

want to see how the whole process works? 👀 i put together a step-by-step guide using Hurry Up! as an example, from the initial scene description and character references to the storyboard, finished film, and final edits. check out the attached PDF to see how you can bring your own short film to life with OpenArt director mode! 🎬

link to the guide: https://drive.google.com/file/d/1ZWzG4kZ43IZnBdR-qPnwv7fozs8N719b/view?usp=sharing

Prompt

Create a 30-second ultra-realistic early-2000s Japanese home vlog featuring the same young Japanese girl with consistent face, hairstyle, and outfit.

Scene 1: She enters her bedroom and finds messy futon bedding, twisted blankets, and scattered pillows.

Scene 2: She removes the pillows, folds the blanket neatly, then folds and rolls the futon mattress and puts it away.

Scene 3: She removes the old pillow covers, replaces them with clean ones, fluffs the pillows, and arranges them neatly.

Scene 4: She sweeps the entire bedroom floor, including corners, then wipes away the remaining dust with a cloth.

Scene 5: She straightens the room, opens the curtains, and admires her clean, peaceful bedroom with a satisfied smile. End with a clear wide shot revealing the finished room.

Keep every scene in exact order. Show each action completely with no skipped steps, sudden transitions, or magically moving objects. Maintain consistent props and room layout. Use authentic 2003 Japanese furniture and handheld MiniDV footage with natural lighting, slight camera shake, autofocus shifts, and mild digital noise. Realistic movements and a satisfying final reveal.

For today’s OpenArt Daily Challenge, I wanted to turn a toddler’s imagination into a full cinematic battle.

In his world, Sai is a battle-worn knight fighting a massive mechanical beast. In reality, he’s just holding his little sword-wielding doll against the family robot vacuum. 😂

The reality reveal was the most fun part to create. Made with OpenArt.

Did the ending work for you?

🎃 One spin can change everything… but some things should never be revealed. 🌀🩸

Welcome to THE FANVERSE. Dare to watch until the end? 👁️

#TheFanverse #HalloweenHorror #AIHorror #AIFilm #OpenArt

A moment freezes. Then it happens again, and it's worse.

Morning in the studio. Dad lifts a full water jug, slips on the wet floor, and time stops. The camera circles him, and from behind it reveals what he can't see: his co-worker frozen in the doorway with a surprise birthday cake, candles lit.

Time resumes. Dad lands soaked with the jug in his lap. The kid rushes in to help, forgetting the cake in his hands... and steps straight into the puddle Dad just made.

Freeze number two. The cake hangs in midair, candles still burning, inches from Dad's face.

So... what happens next? 👇

Made with OpenArt + Seedance 2.5. Both characters are my own originals.

Idea + key prompt in the comments.

i ran FLUX 3 Image at all four resolutions, changed a garment, one word on a cover line and a brand name on a can, and then built one editorial photograph out of ten separate references.

first, the question in the title. is it the most realistic ai image generator?
i cannot answer that honestly. nobody has benchmarked these against each other, and i am not going to call it from a vibe. what i can do is show you what it looks like at pixel level and let you decide.

Four things, each with a proof instead of a claim.

1. The resolution is real, and i checked it at pixel level.

same prompt, four tiers. the delivered files are 1024x576, 1360x768, 2736x1536 and 5456x3072. that top one is 16.76 megapixels, which matches black forest labs' own "about 16 megapixels" for the 4K setting.

then i cropped 512 native pixels from each, with no scaling at all, and put them side by side. at 768p and 1K that 512 pixel window holds the entire head, and there is no pore structure in it anywhere. at 2K the same window holds the eye and brow, and pores start to appear. at 4K the same window holds a single eye, with individual pores, fine vellus hairs, skin flakes, the wet lower eyelid and iris fibres.

same pixel budget. a whole head at 768p, one eye at 4K.

one thing i did not expect. i cropped a patch of bare plaster wall from the 4K and from the 2K, and enlarged both 2x with nearest neighbour so nothing could be invented by interpolation. the 2K wall is clean. the 4K wall carries fine even grain across the whole patch. bare plaster has no texture of its own, so that is noise the model added, and it is the thing that finally stopped looking like a render to me. that is one prompt and one region per tier, so take it as a good sign, not a law.

2. The edits leave the rest alone, and i measured that too.

the headline is that it edits "without changing any other pixel". their own gallery is more honest than that headline: it publishes the percentage of pixels that come back identical per example, and the range across their nine examples is 67.8 percent to 94.1 percent. so even they are not claiming 100.

i measured mine the same way, per channel, at a tolerance of 8 out of 255.

i changed a wool sweater into a waxed canvas coat. 75 percent of the frame came back identical. the mask shows where the change actually went: the entire head, the face, the eyes, the window and the wall are untouched, and the white is the coat and nothing else.

then i changed one word on a cover line, and one brand name on a product can. 99 percent identical, both. the can is the interesting one, because the name is printed on a curved surface, so the type wraps and foreshortens at both edges, and it held anyway.

so the number is not a fixed property of this model. it tracks how much you ask it to change. a whole garment costs you a quarter of the frame. one word costs you one percent. and their published range sits inside my two extremes, which is exactly what you would expect if their examples are bigger edits than a one-word change.

3. The text editing is the part you will actually use.

a brand name on packaging, a cover line, an exhibition date. that is the real job, not a trick.

and the technique is worth knowing, because it is not a mask and not a box. it is naming the things you refuse to lose, one at a time. i was writing "keep the background" and it did not hold. listing the window, the wall, the light and the framing does.

4. Ten references, one frame.

built one editorial photograph out of ten separate references: a face, a jacket, a knit, an earring, a room, a light quality, a wool weave, a palette, a pose and a ceramic vessel.

nine of the ten are visibly doing work, and the one that matters most is the face. the same woman comes back, and she is not a copy of the reference photo, she is a different photograph of the same person. that is what consistency actually means, and it is the thing i was least sure the model could do.

the tenth one failed, and the failure is the useful part. i scoped the wool weave reference to the knit. the weave went onto the jacket instead, and the knit stayed a fine knit. so: a texture reference will fight the object you scoped it to, and the largest nearby object usually wins. if texture placement matters to you, give it its own object or leave it out.

i have not seen that written down anywhere, so i am writing it down.

1. medium first. open with what kind of photograph it is, then the subject. "a close-up documentary portrait photograph, 85mm lens" and then the person. this one contradicts most advice you will read, including the advice on some model pages, which says subject first. every one of black forest labs' own launch captions opens with the medium, and once i switched the look stopped drifting.

2. say what you want, not what you don't. "a deserted beach at dawn", never "a beach, no people". there is no negative prompt field on this model, so anything you describe as missing is just one more thing you described.

3. short is fine, and it gets expanded for you. write more only for what you actually want to control: position, light, text. past that you are diluting the thing you cared about.

4. name the light instead of gesturing at it. "split lighting at ninety degrees", "rembrandt, key at 45", "soft window light from the left". "good lighting" is not a direction, it is a wish.

5. hex for anything that has to be exact, and tie it to one named object. walls painted #0047AB (cobalt blue). the model needs to know which thing you mean.

6. concrete beats praise. "soft window light from the left, 28mm lens, fine film grain" does more than "stunning". if you could not point at it in the frame, cut it.

all six are pulled from black forest labs' own prompting guide for this model, not invented, and the medium-first point in particular is theirs.

what are you trying to get FLUX 3 Image to do that it will not do? drop it in the comments and i will test it on my next run.

(all of it on Openart, if you want to rerun any of it)

It doesn’t just print on paper. It brings paper to life.
Created for the OpenArt Ad Awards, hope you guys like it.

Some moments deserve bullet time. A dad's pride getting destroyed by chili sambal is one of them.

The setup: a young guy dares a 50-year-old dad to eat the hottest sambal at a roadside chicken noodle stall in Jakarta. Dad doesn't back down. Dad never backs down.

Then the sneeze happens.

Time drops to ~2% speed and the camera glides through the frozen moment: noodle strands glossy with chili oil, broth droplets hanging like glass beads, the meatballs, the wonton, his beanie floating above his head, and the kid mid-recoil with his phone still recording. Then time snaps back, and someone has to pay for it. (It's never the kid.)

Made with OpenArt + Seedance 2.5 · 15s · one continuous take.

Idea + key prompt in the comments 👇

Today’s Bad Idea: A superhero whose only power is… finding parking. 🚗💀

Your mission? Take the most ordinary superpower imaginable and make it feel like the fate of the city depends on it.

Think dramatic entrances, packed parking lots, desperate drivers, impossible parking spots, people losing hope… and one hero who somehow always knows where the empty spot is. 🅿️🦸

🎬 Create your video

📤 Post it in the Day 5 challenge thread

🏆 Earn Challenge Points

🎁 Stack enough points to unlock FREE OpenArt credits!

💬 Winner will be announced in the comments!

No complicated rules. Just take the idea and make it your own.

Would you trust this hero to find your parking spot? 👀

A dragon. A sneeze. And suddenly, everything stops

Made with Seedance 2.5 on OpenArt