What if movies were not viewed but experienced?
What if you could star in your own movie?
What if you could change the ending?
https://t.co/SJMAPRkf3t
- Choose your genre
- Cast yourself and your friends
- Change the storyline
Create your own adventure. All in 15 minutes.
Introducing Koyal Experiences.
Generate for free now at https://t.co/SJMAPRkf3t
@krea_ai's Krea 3 Edit is one of the strongest models on the board at precise realistic edits.
It ranks 2nd on realistic adding (inserting an object into a scene), with a 56.8% win rate. Its other leading placements:
- 2nd, realistic color (shifting the color of a specific element), 54.6%
- 3rd, realistic text editing (replacing text in-frame), 55.9%
- 6th, realistic removing (taking an object out and rebuilding what was behind it), 53.9%
- 6th, animated color, 51.6%
Color palettes is its weakest edit in both styles, 15th animated and 13th realistic. Regrading the whole image is where it falls behind. Worth knowing if color grading is your use case.
Every one of its top results is a targeted edit. Adding an object, recoloring one element, replacing a line of text and removing an object all change one region of a photograph and leave the rest alone. The hard part is the boundary: the new object has to cast the right shadow, and the new text has to follow the surface it sits on. Krea 3 Edit places top three on three of these edits in realistic, and no model measured adds an object to a realistic scene more convincingly, apart from one.
Measured over 65k blind pairwise votes.
https://t.co/dsWazS1s3u
@recraftai's Recraft V3 makes its Film Arena debut on the image editing board.
Its strongest result is animated expression (changing a character's facial performance), where it wins 50.7% of its matchups. Its other leading placements:
- 11th, animated color palettes (regrading the whole image to a new palette), 46.6%
- 15th, realistic color palettes, 43.2%
Realistic editing is its weakest area, and removing, text and lighting rank at the bottom of the board. Worth knowing if realistic photo editing is your use case.
Its expression result is the one to watch. Changing a character's face is one of the hardest animated edits, because the new expression has to fit the character's existing style and proportions while everything else in the frame stays fixed. Here, Recraft V3 wins more of those matchups than it loses, which gives the next version a clear starting point.
Measured over 45k blind pairwise votes.
https://t.co/dsWazS1s3u
@bfl_ai's FLUX 3 Max enters the Film Arena image-to-video board with its strongest results in realistic video.
It ranks 12th on realistic object handling (props and objects holding their shape across frames), with a 52.4% win rate. Its other leading placements:
- 19th, realistic high movement (fast action), 50.9%
- 21st, realistic multi-character (two or more subjects interacting in frame), 50.6%
- 26th, realistic emotions (expressive gesture and facial performance), 48.1%
Object handling is where video models most often fail, and the failure is easy to see. A cup changes shape as a hand picks it up, or a door handle disappears when the camera moves. FLUX 3 Max wins most of its realistic object handling matchups. It also stays above a 50% win rate on high movement and multi-character, the two categories where motion is most likely to break that coherence.
Measured over 20k blind pairwise votes.
https://t.co/dsWazS1s3u
@bfl_ai's FLUX 3 Edit produces the best looking realistic restyles on the Film Arena image editing board.
It ranks top seven on five editing benchmarks:
- 4th, realistic background (replacing the scene behind a subject), 55.8%
- 5th, realistic text editing (replacing text in-frame), 54.7%
- 5th, animated color palettes (regrading the whole image), 53.2%
- 6th, realistic character swap (replacing one subject with another), 54.4%
Where FLUX 3 Edit leads is image quality. It ranks 1st on visual quality for realistic style transfer at 64.6%, and 3rd on visual quality for both realistic background and lighting, at 58.9% and 61.6%. These edits change most of the frame, and the result still has to look photographed: light has to fall consistently and surfaces have to hold together.
Measured over 68k blind pairwise votes.
https://t.co/dsWazS1s3u
@ideogram_ai's Ideogram 4.5 model is one of the strongest models on the board at removal and seamless blending.
It ranks 3rd on animated removing (taking an object out and rebuilding what was behind it), with a 53.0% win rate. Its other leading placements:
4th, realistic color (shifting the color of a specific element), 52.4%
5th, animated color, 51.9%
12th, realistic text editing (replacing text in-frame), 52.5%
Animated lighting and animated background sit near the bottom of the board. Edits that repaint the whole frame are its weakest area. Worth knowing if relighting or scene replacement is your use case.
Its strongest result is blending. It ranks 2nd on blending for realistic removing at 57.0%, and 3rd on blending for realistic expression at 60.7%. In both edits the model changes one region and has to leave everything around it untouched, which is where most models show a visible boundary. Ideogram 4.5 places in the top three on blending for both, even where its overall rank is in the middle of the board.
Measured over 70k blind pairwise votes.
https://t.co/dsWazS1s3u
There is a new state of the art in animated object removal.
@Google's Nano Banana 2.1 ranks 1st on animated removing (taking an object out and rebuilding what was behind it) on the Film Arena image editing board, with a 54.8% win rate. Its other leading placements:
2nd, animated expression (changing a character's facial performance), 60.3%
3rd, animated color palettes (regrading the whole image to a new palette), 56.1%
3rd, realistic color palettes, 55.6%
4th, realistic adding (inserting an object into a scene), 56.0%
Realistic removing is 10th, well below its animated result. Worth knowing if live action cleanup is your use case.
Blending decides whether an edit reads as part of the original image or as something added afterwards. Most models follow the instruction and still leave a visible seam where the edit meets the frame. Nano Banana 2.1 ranks 1st for keeping identity intact on character swap in both styles (replacing one subject while keeping who they are), and 2nd on blending across five editing types. It leaves the fewest seams of any model measured.
Measured over 75k blind pairwise votes. No self-reported evaluations.
https://t.co/dsWazS1s3u
Introducing Koyal Act 3: The Agentic Studio for Filmmaking.
It is the first end-to-end agentic system for feature-length movie production.
Filmmakers with it can create, direct, and iterate on long form AI content from scratch.
Coming out of beta soon.
SINGLE UNTIL SERIES B for your 2027 GRAMMY®️ consideration in the Record of the Year, Song of the Year, Best Pop Duo/Group Performance, Best Music Video categories.
Thank you for all the love.
#ForYourConsideration#Grammys#RecordingAcademy#FYC
Today, we’re releasing the first official anthem to celebrate being a single founder.
Society has a million songs celebrating love, marriage, and heartbreak. But somehow, nobody thought being single was worth celebrating.
Until now.
A few months ago, @contextconor started a movement: Single Until Series B.
He recruited me early. I joined his crusade, at great cost to my public reputation.
Today, we’re taking it a step further with an official anthem and music video for single founders everywhere.
Inspired by the very real suffering of pre-PMF founders, Conor composed the song, enhanced with Suno.
We shot the video in an hour with friends, then expanded it into a full hybrid production with our team at @KoyalAI using https://t.co/fL464fcPs6.
Some people find love. Others find product-market fit.
This one’s for everyone who’s found neither.
spoke at uncanny india in delhi last week, @KoyalAI film fest and their first one in the country with the gov of india. the panel was titled “production rewritten” and we talked about ai
every generation of filmmakers gets a new technology thrown at them and has to decide whether to accept it, reject it, or reason with it. this is ours. it’s not about whether you use it anymore, it’s how
we made 5 feature-length animated films this year doing r&d. never set out to make ai films. just films. now that’s compounding and we’re onto a moonshot. with genai, long-form is the only frontier worth fighting for. i hope that in a few years and there won’t be ai filmmakers. just filmmakers
we presented machhi mafia, our upcoming AI-animated series, at uncanny valley fest delhi on saturday. a crime thriller set in mumbai, centred around mafia, koli women, life at sea and land. magic realism, borrowing from marquez and ben okri
my brother and co-founder @anupamsidhant was on the panel and jury too. speaks fucking well, this kid.
thanks @meh_agarwal@KoyalAI 🐟
we presented machhi mafia, our upcoming AI-animated series, at uncanny valley fest delhi on saturday. a crime thriller set in mumbai, centred around mafia, koli women, life at sea and land. magic realism, borrowing from marquez and ben okri
my brother and co-founder @anupamsidhant was on the panel and jury too. speaks fucking well, this kid.
thanks @meh_agarwal@KoyalAI 🐟