Umesh's FLUX 3 park film is flawless until a single leaf lands
Umesh's 15-second downtown park film holds one nostalgic, film-stock grade across nine shots, animated with FLUX 3 from a Midjourney v8.2 still he made with a public style code. The only giveaway is a dry leaf that lands on a scooter seat with no bounce. He hasn't said which platform, prompt or settings made the video.
Low morning sun slides across a downtown park. An old tree stands in the middle of the lawn with a row of scooters parked in its shade, palms line the far side, and glass towers rise behind everything. The camera drifts up the trunk, tilts into the canopy and settles on the scooters, where one dry brown leaf drops onto a seat and stays there. Only in the last two shots does anyone step into the story: an old man and a small girl in a yellow dress, walking hand in hand past the tree and away down the path.
The 15-second film was posted on X on Monday 12 October by Umesh (@umesh_ai), a creator, filmmaker and developer with about 48,000 followers who is also an Adobe Firefly Ambassador. His caption is a single line: "Flux 3 is awesome at keeping the base image style consistent through out." It is our Video of the Day, and on that claim he is right.
A park with a film-stock soul
These are wonderfully graded shots of a city-center park, and they carry a nostalgic, film-stock quality: warm low light, soft grain, lifted shadows and greens that sit slightly faded. The look holds across all nine shots, from the opening wide to the closing walk away, which is exactly the consistency his caption is about.
The subject is the first signs of autumn, and the film is restrained about it. The canopy is still mostly green, palms and ferns included, and the turn of the season arrives as that one dry leaf on the scooter seat.
We don't meet anyone until the closing shots, about 11 seconds in, and everything about the two characters is wonderfully human. She looks up at him as they walk, he keeps an unhurried pace, and at 1080p there is no warping in faces or hands to break the spell.
The one giveaway
The only minor giveaway that this is generative film is weight. The leaf falls in about five frames and lands on the seat with no bounce, no slide and no flutter, then sits dead still, as if it weighed far more than a leaf does. It is a tiny moment just before the six-second mark, and it is worth watching for. Apart from that leaf, the film is flawless.
That is the kind of detail physics benchmarks try to measure. Black Forest Labs, the company behind FLUX 3, says the model tops Google DeepMind's Physics-IQ test, which scores video models on real-world physics; that is the company's claim, and we have not checked it.
How he made it
The recipe sits in his own posts. Twenty-eight minutes before the video, Umesh posted four Midjourney stills made with a style reference code, or sref, a numeric shortcut that applies one consistent look to every image: --sref 3534013297 at a style weight of 500 and stylize 600, on Midjourney v8.2. He put the full prompts in the images' alt text. The first still, "A peaceful urban park surrounded by towering glass skyscrapers, an ancient tree, electric bicycles resting beside a pathway", is the opening frame of the film, matched composition for composition.
- Step: Look. Tool: Midjourney v8.2, --sref 3534013297, --sw 500, --stylize 600; What the posts show: Four stills posted at 06:48 London time on 12 October, prompts in the alt text
- Step: Start frame. Tool: Midjourney still 1 (the park); What the posts show: The film opens on this still, matched by composition
- Step: Motion and cuts. Tool: FLUX 3, per his caption; What the posts show: No prompt, platform, settings or number of generations given
- Step: Edit. Tool: Not stated; What the posts show: Nine shots in 15 seconds; one generation or several cut together is unknown
FLUX 3 is Black Forest Labs' multimodal model, announced on 23 July and trained on images, video and audio together. Its video version became generally available on 4 August through the company's API, the interface that developers and platforms plug into, and through partners such as fal and Higgsfield. It turns a prompt plus one to ten images into clips of 5 to 20 seconds, and the model page lists several scenes per generation. At the list price of $0.29 a second for Full HD, a 15-second clip like this costs about $4.35 on Black Forest Labs' own API, though Umesh hasn't said where he ran it.
What he hasn't said matters as much. The caption names only "Flux 3", with no prompt and no word on whether the nine shots came from one generation or several cut together. Two people asked about his process in the replies and neither got a workflow answer, so treat this as a FLUX 3 film that starts from a Midjourney still, not as proof of one prompt and one take. This piece makes no judgment on the soundtrack.
In practice, for anyone making work this week, the useful part is the order of operations. Lock the look in a still first with a style code you can reuse, hand that still to a video model as the start frame, and judge the result on whether the grade survives the cuts. Umesh ran another still from the same set through Seedance 2.5 on Runway that morning, so his feed doubles as a rough side-by-side for anyone comparing models.
Who made it
Umesh posts every day, mostly prompt shares for Seedance 2.5, Kling and Midjourney style codes, with plenty of optimism about AI tools mixed in. This film reached about 11,600 people by Tuesday morning, roughly four times his usual reach and his most-liked post of the week, and 189 people bookmarked it, which suggests they are saving the recipe. He is an Adobe Firefly Ambassador, and FLUX models are available inside some Adobe products, but this post doesn't mention Adobe, carries no paid-partnership label and has no watermark.
Watch the film on Umesh's X post at x.com/umesh_ai/status/2109528683316961404, and find the Midjourney stills and their prompts at x.com/umesh_ai/status/2109521650316358055.
Film and frames by Umesh (@umesh_ai), made with FLUX 3 from a Midjourney v8.2 still and shared on X. They are used here with credit. If you are the maker and want any of it taken down, message @the__generated on X and it comes down.
Your Voice
Style drift is my biggest headache for multi-shot scenes. Did you have to lock anything (seed / reference strength) to get this consistency, or is Flux 3 just that good out of the box?
Flux 3 keeping the style steady is a nice win.
Base image style consistency across frames is the hard part.
Quoted from public posts on X and credited by handle. If one of these is yours and you want it removed, message @the__generated on X and it comes down.
Got something to say? Tag @the__generated on X