Star in your own movie today at xp.koyal.ai - kookoo

INDIA'S BIGGEST AI FILM HACKATHON. OCTOBER 10TH NEW DELHI. REGISTER NOW: luma.com/r6qve8rn KOOKOO

No model on the image editing board makes a cleaner edit than @Alibaba_Qwen's Qwen Image 2.1 does on animated removing. It takes 1st of 29 on edit accuracy for animated removing (taking an object out and reconstructing what was behind it), the single criterion that measures whether the instruction was actually followed. Its strongest placements overall: - 5th, animated adding (inserting an object into a scene) at 53.4% - 6th, realistic text editing (replacing text in-frame) at 53.5% - 6th, animated text editing at 52.8% - 7th, realistic removing at 53.1% - 9th, realistic expressions (changing a facial performance) at 52.9% Lighting is 27th animated and 24th realistic, its worst result by a wide margin, and background editing is 20th and 21st. Both are global edits that repaint the whole frame. Worth knowing if relighting is central to your pipeline. The pattern across its wins is that every one of them is a local edit: one object added, one object removed, one line of text replaced, one expression changed, with the rest of the frame expected to survive untouched. That is the harder half of editing to get right, because a model that redraws generously scores well on visual quality and badly on everything the user actually asked for. Qwen 2.1 averages 7.9th on semantic consistency across both styles, its best criterion of the five, and it holds that while placing top eight on four separate local-edit attributes. Measured over 65k blind pairwise votes. filmarena.ai

Being an underdog is our culture

Animated camera movement has a new leader. @bfl_ai's FLUX Edit model takes 1st on animated camera movement (adding or altering a camera move) at 57.6%, its widest margin on the board. Behind it: - 2nd, realistic expressions (changing a facial performance) at 53.6% - 3rd, realistic removing (taking an object out and reconstructing what was behind it) at 51.2% - 4th, realistic pace (retiming a shot without breaking its motion) at 52.4% - 5th, animated removing at 49.8% Relighting a shot for a different hour and replacing the environment around a subject are where this model places last. Worth knowing if either is central to your pipeline. Camera movement and pace are the two attributes that ask a model to change how a shot moves while leaving what is in it alone. Most editors couple the two, and re-timing or re-framing drags the subject with it. FLUX Edit places 1st and 5th on those two in animated, and 3rd combined on camera movement across both styles, which is a narrower specialty than the leaders hold but a cleaner one. Measured over 35k blind pairwise votes. filmarena.ai

KOOKOO

JOIN US FOR THE BIGGEST CREATIVE ML HACKATHON AT CMU: luma.com/gdw75ze5

There is a new state of the art in image editing. @OpenAI's @ChatGPT Image 2.5 Sunburst is 1st of 28 on both realistic and animated, the first model to lead both boards at once. Its first places: - 1st, realistic lighting (relighting a scene in place) at 61.9%, its widest margin anywhere - 1st, animated expression (facial performance and gesture) at 61.6% - 1st, realistic background (replacing the environment around a subject) at 60.1% - 1st, realistic text editing (changing text in an image without redrawing what surrounds it) at 58.1% - 1st, animated text editing at 57.0% Realistic removing (taking an object out and reconstructing what was behind it) is 3rd, its weakest placement. Animated adding is 6th. Both are local edits, and both are where the model gives ground. The pattern in the wins is scene-level control. Lighting, background and expression each change something that touches the whole frame, and the standard failure is everything else drifting with it: the subject shifts when the light does, the face changes identity when the expression does. Sunburst holds the rest of the image still in all three, and does it in both styles, which no other model on the board manages. Measured over blind pairwise votes across 28 models. filmarena.ai

Muse Image 1.0 by @Meta takes first place on five separate editing attributes. 3rd of 24 on realistic and 2nd of 24 on animated, with the top spot on: - 1st, animated background (replacing the environment around a subject) at 61.5% - 1st, realistic swap character (substituting the subject in a shot) at 59.3% - 1st, realistic adding (inserting a new object into a scene) at 58.3% - 1st, realistic text editing (changing text without redrawing what surrounds it) at 56.8% - 1st, animated text editing at 56.6% Realistic expression (facial performance) is 9th, well off its animated 2nd. MAI Image 2.5 leads that category at 59.9%. Worth knowing if photoreal faces are your use case. Text editing is the result that carries. It is first in both styles, and it is the attribute where models most often succeed at the edit and fail at everything around it: the sign warps, the surface relights, the surrounding detail quietly redraws. Holding both the new text and the untouched remainder of the frame is a different problem from either alone, and Muse is the only model that leads it on realistic and animated at once. Measured over blind pairwise votes across 24 models. filmarena.ai Congrats to team to the Muse team at @AIatMeta!!

I'm in a mortal kombat inspired video game?? - KOOOKOOOOOO (with the intention of beating you up)

realtime AI dating simulator only on koyal - kookoo

chopped.koyal.ai/date is powered by H3 Max by @fal, the top open source model on filmarena.ai

go checkout SF's Chopped List and on virtual dates with sf's most chopped founders at chopped.koyal.ai :)