How did GPT-6 end up working for ByteDance's Seedance??
And this hit comes from someone juggling multiple roles—
Write stories, set scenes, cast actors, position cameras, create storyboards, generate videos, and edit the final cut…
Alright, alright, this time GPT-6 is directing, and Seedance 2.5 is handling production.
At first glance, two anime characters are drawing their swords and fighting in a Japanese alleyway.
The camera sweeps briefly over the eaves, then rapidly pushes in close to the characters; the characters transition from standoff to sprinting to close-quarters combat, yet all these shots remain within the same scene.
In today’s AI video scene, just looking at the final product might not be enough to knock you off your seat (doge).
But look again at the task Higgsfield assigned to Astra:
Create a sword-fighting story, pre-visualize it in Blender, generate shots using Seedance 2.5, and edit the footage into the final version while maintaining consistency in characters and scenes.
In other words, humans provided only a high-level requirement, and GPT-6 Astra handled everything from the brief to the final output entirely on its own.
No, at the GPT-6 launch, they were still designing circuit boards in KiCad and building houses in Blender—
It's only been two days, and netizens have already casually assigned it to a film studio...
GPT-6 as director, Seedance handles the performance
The entire implementation path looks somewhat like an AI-powered animation production team.
After receiving the task, GPT-6 Astra must first create an entirely original story suitable for filming.
Who is fighting whom, why they are fighting, where the fight is taking place, how the actions unfold, and how many shots are needed—these things must be clarified first.
The story is set; now it opens Blender.
First, construct a relatively simplified 3D scene with streets, buildings, characters, and props according to the storyline, then place the two characters within it and define their positions and movement paths across different shots.
What is being done in this stage is film previsualization.
After all that back-and-forth, the entire fight sequence had already been rehearsed in Blender using 3D models.
GPT-6 Astra then generates reference frames corresponding to different shots based on the previsualization footage, and passes the character designs, scene visuals, and motion references to Seedance 2.5.
At this point, ByteSeedance officially joins the team.
It transforms static reference frames and rough storyboards into final videos with realistic character animations, camera movements, and an anime-style aesthetic.
Of course, after generation was complete, Astra was still at work.
It then drags the segments generated by Seedance into an editing tool, selects shots, arranges them on the timeline, and assembles them into a final version before exporting the completed video.
More precisely, GPT-6 assembled multiple tools and models into a temporary set:
GPT-6 Astra writes the story, operates Blender to complete the previsualization, then calls Seedance 2.5 to generate the final shots, and finally returns to the desktop to complete editing and export.
People are waiting outside to collect their coins...
Putting AGI aside for now, Seedance 2.5 has definitely handled this well.
Next, netizens took this玩法 to the Leaf Village.
Astra first creates a 3D blockout, retaining only the building structures, character positions, and approximate spatial relationships in a simplified model.
Then, set up six cameras in advance in Blender, each determining how one of the six shots will be filmed.
After the shot is finalized, a frame is exported at each cut point.
These frames are then refined by GPT Image 2, transforming the original rough 3D grayscale models into keyframes with a Naruto-style aesthetic.
Finally, Seedance 2.5 debuted—a 30-second fan-made short film of Konoha Village, composed of six scenes.
They haven’t had enough of animations—GPT-6 and Seedance are now targeting Chinese short dramas.
This round’s concept is more abstract: in one sentence, adapt a 30-episode Chinese short drama into an American version.
In the comparison video, the main storyline and shot pacing are nearly identical, but the characters, settings, language, and cultural context have all been completely changed.
The Chinese AI actor's face was replaced with that of an American AI actor, while the original room, street, and clothing were regenerated, with the framing and character movements closely following the original footage.
GPT-6, Fable5.1, and Gemini 3.8 have all come to Seedance to film.
Moreover, GPT-6 isn’t the only one coming to Seedance to form a team.
Over the past two days, GPT-6 Astra, Fable 5.1, and Gemini 3.8 Flash alternately generated prompts, which were then uniformly turned into videos by Seedance 2.5.
Alright, alright, alright—Seedance, go live!!!
The same brief produced three distinctly different results when handed to different models.
Some versions focus on character actions, with the camera rapidly advancing alongside the events; others emphasize composition, lighting, and atmosphere, giving the overall feel of a meticulously story-boarded sequence; still others strictly follow the task at hand, delivering complete information but with a less pronounced directorial style.
GPT-6 Astra and Fable 5.1 were still subjected to multiple rounds of testing, remaining at Seedance 2.5.
For example, hand-drawn Japanese ghost stories:
Alternatively, ink wash animation:
Forget about who wins between GPT-6 and Fable—Seedance’s ink wash effect is truly impressive.
After comparing the three large models, the winner hasn’t been settled yet—but ByteDance’s Seedance has already taken center stage...
After Seedance clocked out, GPT-6 went to work as a colorist at DaVinci.
After finishing work on set, GPT-6 immediately took on a second job as a colorist.
This time, GPT-6 opened DaVinci Resolve.
An internet user provided GPT-6 Astra with a reference image, asking it to adjust the colors of the original video to match a similar style, as a professional colorist would.
It requires observing the color temperature, contrast, shadows, highlights, and overall color tone in the reference image, then adjusting the color grading tools in DaVinci. Click here Edit the video.
The result after 4 minutes:
Compared to the original, the video appeared flat and gray overall; after processing, the colors are significantly more vibrant, shadows are darker, and the warm-cool balance is closer to the reference image.
Of course, netizens' opinions on this AI colorist are not as consistent as those on the animated short film.
Some people find it convenient that the reference image style matching is completed in just four minutes;
Some have bluntly said that after the color grading, it actually looked worse... Color grading should serve the story; color itself can be a narrative motif—AI clearly still cannot truly understand these creative intentions...
It can only be said that this is a classic tradition of the film and television industry—before the tools have even been brought on set, the debates over aesthetics have already begun.
At the launch event, OpenAI was still working to demonstrate that Astra could operate various professional software; by the time it reached netizens, the instructions had been creatively adapted in countless ways.
As for Seedance—
Sitting at ByteDance, GPT-6 somehow became my best partner.
Sora: What about me???
Reference links: [1] https://x.com/higgsfield_ai/status/2096352124754104618?s=20 [2] https://x.com/groovestreetgen/status/2096354475879600432?s=20 [3] https://x.com/Pinkuo_/status/2095754189548986426 [4] https://x.com/eachlabs/status/2095911062499381467 [5] https://x.com/eachlabs/status/2095911062499381467 [6] https://x.com/higgsfield_ai/status/2095609438946394380 [7] https://x.com/AngryTomtweets/status/2095654817154908253 [8] https://x.com/higgsfield_ai/status/2096334648016253339
This article is from the WeChat official account "Quantum Bit," authored by Wen Le.