Skip to content
AI Primer
workflow

Ozan Sihay shares ten first-attempt FLUX 3 Max video tests at 720p

Ozan Sihay shares first-attempt results from ten FLUX 3 Max video tests without selecting the best takes. Each uses 15-second, 720p text-to-video generation at high reasoning to test constraints such as physics and identity.

4 min read
Ozan Sihay shares ten first-attempt FLUX 3 Max video tests at 720p
Ozan Sihay shares ten first-attempt FLUX 3 Max video tests at 720p

TL;DR

One mirror test asks a reflection to stop following its owner, look at the camera, then return to sync. An origami sequence specifies paper, glowing particles and feathers as three readable stages. FLUX 3 Max is fal Research’s post-trained video model, built with Black Forest Labs on FLUX 3.

Generation controls

Ozan Sihay published a full English shot brief beneath each Turkish test description. The prompts are the keeper: they expose the requested action and constraints alongside each generation.

For context, fal’s API reference documents these controls:

  • Duration: 5 to 20 seconds.
  • Resolution: 720p or 1080p.
  • Reasoning: low or high, with high as the default. Higher effort slows generation.
  • Audio: generated by default, with an option to disable it.

Sihay gives each test an objective without assigning pass/fail scores. The descriptions below summarize the requested behavior.

Physics, fingers and parkour

  • Test 1, chain reaction: A marble must trigger dominoes, a lever, a pendulum, a broken liquid vial, a waterwheel, gears and finally a bulb. The brief requires a continuous macro shot with causal order and momentum transfer preserved.
  • Test 3, watchmaking: The watchmaker prompt requires tweezers to place a gear on a jewel bearing, followed by a screwdriver tightening a screw. Fingers must remain anatomically correct, hands separate and tools rigid during a slow 100mm macro orbit.
  • Test 9, parkour: The athlete must sprint, vault an air-conditioning unit, jump a three-meter gap, roll, wall-run and backflip. The brief specifies believable foot contact, momentum-driven clothing and a tracking chase shot with a speed ramp at landing.

Turkish lettering and dialogue

  • Test 2, typography: Two exact strings, “DUR, BİR BAK!” and “ÇAY 15 TL,” must remain legible for the full clip. The scene combines neon, chalk lettering, wet-street reflections and a focus pull from the sign to a man’s face.
  • Test 7, dialogue: A tea-house conversation assigns three speaking turns: older man, young woman, older man. The prompt requests synchronized Turkish speech, natural expressions, eyes following the speaker and stable hands around the tea glasses.

Crowds, camera paths and waterlines

  • Test 4, identity through occlusion: In the Grand Bazaar test, a woman’s face, hair, red coat and silver scarf must remain consistent as people cross between her and the camera. Background faces must also stay stable.
  • Test 5, continuous geography: The camera path runs through a keyhole, along a hallway, down stairs, through a cellar window, into a snowy garden and above a village. The prompt explicitly prohibits cuts, dissolves and spatial morphing, while requiring consistent scale and parallax.
  • Test 6, split-level water: A woman dives from a pier in a half-underwater composition. The brief requires a continuous waterline, coherent above-and-below geometry, refraction, caustic light and believable submerged fabric.

Paper transformations and mirror geometry

  • Test 8, transformation: Beyond the three material stages, the camera must move from a locked-off medium macro shot into a tracking follow as the living crane flies through a window toward a mountain valley at sunrise.
  • Test 10, mirrors: The surrounding geometry must support recursive reflections between opposing mirrors, correct left-right reversal and glass that never reveals the camera, alongside the deliberately disobedient reflection.

A separate Claude animation

Sihay also shared a separate 15-second stylized Farrah Fawcett portrait animation, crediting Claude Opus 5.5 (High) in the caption. It sits outside the ten-test FLUX thread.

Free generations and paid access

Sihay initially advertised ten free 15-second, 720p generations through fal, then said most respondents reported 50. fal’s free-video page, updated October 21, lists five free generations every 24 hours for signed-in accounts and says account activation requires a card on file.

For paid use, fal’s FLUX 3 Max page lists a 75% discount through November 7: $0.0425 per output second at 720p, against a $0.17 list rate. At the discounted rate, a 15-second clip costs about $0.64.

Further reading

Discussion across the web

Where this story is being discussed, in original context.

On X· 3 threads
Physics, fingers and parkour1 post
Crowds, camera paths and waterlines1 post
Free generations and paid access2 posts
Share on X