verq

video ai researcher in bay area, CA • cinema, art, perception • movies are prompts

[verq @verqious]

AI cinematography series - EP1:

Can AI understand a DUTCH ANGLE?

In cinema, tilting the camera creates psychological unease (think The Third Man).

I tested this with Runway and Grok, paid subs.

Both completely failed. Here's why that matters:

[verq @verqious]

@DaveShapi Not James Cameron. Better.
https://t.co/KhLdAQIy4z

[verq @verqious]

@DynamoSuperX eagerly waiting for avatar 4 and 5!

[verq @verqious]

@c_valenzuelab we're living the visual version - scrolling feeds where half is Seedance 2.0, half is phone footage, and our brains can't correlate which is which anymore.

[verq @verqious]

fair point on consistency, but that's exactly why this feels like an inflection point.
remember when GPT couldn't write coherent paragraphs? the gap between 'impressive clips' and 'full narratives' might be shorter than we think.
imo, even if full films are 5-10 years out, the tooling for pre-vis, concept art, and indie creators is already transformativ.

[verq @verqious]

honestly, people are sleeping on Video AI.

this is just ByteDance Seedance 2.0. imagine what 3.0 does - probably forces a completely new paradigm in moviemaking altogether.

think about it: not much has changed in how we *perceive* cinema. screens got bigger. that's... it?

maybe this is where cinema actually starts changing. maybe the new cinema is spheres? immersive spaces instead of flat rectangles?

especially when the cost to produce these experiences is dropping to near-zero.

the distribution changed. the medium? still stuck in 1960.

that might be about to break.

[verq @verqious]

this is much better than many hollywood scenes!

bytedance seedance 2.0 is cinema

[verq @verqious]

this is much better than many hollywood scenes!

bytedance seedance 2.0 is cinema

[verq @verqious]

this is literally how the next billion-view YouTube music videos are going to be made!

[verq @verqious]

@jm_alexia one-step generation is the unlock everyone's chasing (without distillation).
the real test isn't benchmark numbers - it's whether Drifting Models can handle the same prompt-steering control that makes diffusion useful for film work.
let's hope so.

[verq @verqious]

Damn, this building doesn't need VFX to feel like a sci-fi set. That's the whole point.

Brutalist architecture does what most CGI can't - it creates mood through mass and void, not decoration.

Clean sight lines, natural framing, built-in depth. The camera just shows up and the shot composes itself.

Otani understood cinematography without ever touching a camera.

[Sci-Fi Archives @SciFiArchives]

The Kyoto International Conference Centre designed by Sachio Otani

[verq @verqious]

@CharaspowerAI @Kling_ai INTERESTING: Shot breakdowns used to live in director's heads, then on storyboards, then in Premiere timelines. Now they're tokenized instructions.

[verq @verqious]

@wildmindai Directors been doing this since Eisenstein - shot lists are literally PROMPTS.
Kling wanting rack focus language is not a bug, it's collaborative authorship happening in the latent space.

[verq @verqious]

@c_valenzuelab The Lumière brothers didn't know they were inventing cinema either. They thought they built a better projector.

[verq @verqious]

AI cinematography series - EP7:

I tested 7 cinematography techniques against Grok and Runway.

Final test - CHIAROSCURO: Caravaggio-style lighting. Half the face in light, half swallowed by shadow.

Tutorial: https://t.co/YbypYRinXI

This was the only test about LIGHT, not cameras. And both nailed it.

Runway: cinema-grade detail, skin textures, deep shadows. But temporal flickers on skin broke the illusion.

Grok: rock-solid consistency frame to frame. But the fine details weren't there.

After 7 days, the pattern is clear:

AI understands what a scene LOOKS like.
It doesn't understand how to FILM one.

Lighting, mood, atmosphere - yes.
Camera angles, movement, lens physics - still broken.

These models were trained on frames, not cameras.

Final score (see table attached):
Grok: 5 | Both passed: 1 | Both failed: 1
Runway: 0

This series taught me more about how AI 'sees' than any paper I've read.

#AIVideo #Cinematography #AIFilmmaking #Runway #Grok

[verq @verqious]

@howdyhowdhow @basediogenes @Movies_analyst lol the point is about the constraint forcing different creative choices.

Gallo couldn't review dailies or do safety takes with reversal. that's a different production reality than "we could reshoot but won't."

[verq @verqious]

@InSchlimeOut @UfoBeef @ludoviciochem @PJaccetturo fair. the slop is real.

my bet: people who understand craft will use these tools differently.

[verq @verqious]

@dustinhollywood @nickfloats Higgsfield's whole pitch = "you're the director now"
but the model decides composition, timing, movement.

it's like lottery ticket at this point.

[verq @verqious]

@elonmusk @grok surprisingly grok understood all of the 6 cinematography terms.

runway is dogshit.
https://t.co/733NAUtzxJ

[verq @verqious]

AI cinematography series - EP1:

Can AI understand a DUTCH ANGLE?

In cinema, tilting the camera creates psychological unease (think The Third Man).

I tested this with Runway and Grok, paid subs.

Both completely failed. Here's why that matters:

[verq @verqious]

@NolanAnalyst 57 days but with 6+ months of "pre-production".
People think "fast shoot = rushed"

Nope.

Nolan's secret is frontloading ALL creative decisions into rehearsals + shotlists.

[verq @verqious]

@AlfonsoNation @ParamountPics The problem isn't the tools, it's the "diluting previz window"
Don't constraint artist ruthlessly give them the time & AI tools and make legendary content again.

[verq @verqious]

@ItHowandwas Eight independent motion vectors without collapse. People still fight with me that Video AI is shit.

BTW, The latent space must be doing some insane attention masking.