Bob Alice

tuning out the paywalls. one free open-source tool a day, tested on my own machine. what it replaces + the install line.

[Bob Alice @oscilloo]

GLASS BRIDGE CRACKING UNDER A CROWD

2M likes. 57,000 comments

A packed glass skywalk over a forest gorge. A woman lies on the panels holding a boulder above her head, the glass fractures, and sections give way with people on them

the creator labels it as an AI simulation in the caption, plainly, and it still cleared two million:

→ the setting is a real category of attraction. glass bridges exist, people are already nervous on them, and that fear arrives pre-installed

→ the crowd density does the work. a single person falling is a stunt, a walkway full of people is a systems failure

→ the fracture pattern radiates from the load point and branches unevenly, which is correct for tempered glass and the hardest part to fake

→ the camera stays above and static, which is exactly what a fixed security camera would do

→ and the disclosure line sits in the caption, where nobody reads it, not in the frame

that last point is the part worth arguing about on an account that cares about how these tools get used

the label is real and voluntarily added, which is more than most creators do. but caption text is not part of the video, so every repost, every screen recording and every download strips it instantly, and what circulates is an unlabelled clip of an attraction type that actually exists

this is the gap that provenance standards were built for. C2PA metadata travels with the file, survives re-encoding in principle, and is readable by platforms rather than by humans. it is also almost never enabled by default in consumer tools

so the practical position for anyone making this kind of work: put the disclosure inside the frame, not under it. a corner label costs nothing and is the only part that survives a repost

if you want to test where a viewer's doubt actually sits, image-to-video off one still is the cheapest way in. @Picsart does it from a phone

he labelled it honestly. the label came off the moment somebody else posted it

[Valentin @martynov014]

https://t.co/I1Z0BwhrSi

[Bob Alice @oscilloo]

@martynov014 grandma sheesh

[Bob Alice @oscilloo]

GLASS BRIDGE CRACKING UNDER A CROWD

2M likes. 57,000 comments

A packed glass skywalk over a forest gorge. A woman lies on the panels holding a boulder above her head, the glass fractures, and sections give way with people on them

the creator labels it as an AI simulation in the caption, plainly, and it still cleared two million:

→ the setting is a real category of attraction. glass bridges exist, people are already nervous on them, and that fear arrives pre-installed

→ the crowd density does the work. a single person falling is a stunt, a walkway full of people is a systems failure

→ the fracture pattern radiates from the load point and branches unevenly, which is correct for tempered glass and the hardest part to fake

→ the camera stays above and static, which is exactly what a fixed security camera would do

→ and the disclosure line sits in the caption, where nobody reads it, not in the frame

that last point is the part worth arguing about on an account that cares about how these tools get used

the label is real and voluntarily added, which is more than most creators do. but caption text is not part of the video, so every repost, every screen recording and every download strips it instantly, and what circulates is an unlabelled clip of an attraction type that actually exists

this is the gap that provenance standards were built for. C2PA metadata travels with the file, survives re-encoding in principle, and is readable by platforms rather than by humans. it is also almost never enabled by default in consumer tools

so the practical position for anyone making this kind of work: put the disclosure inside the frame, not under it. a corner label costs nothing and is the only part that survives a repost

if you want to test where a viewer's doubt actually sits, image-to-video off one still is the cheapest way in. @Picsart does it from a phone

he labelled it honestly. the label came off the moment somebody else posted it

[Valentin @martynov014]

https://t.co/I1Z0BwhrSi

[Bob Alice @oscilloo]

@DmitroCP LOOOL BOAT IS DEAD

[Bob Alice @oscilloo]

@martynov014 is it real monkey???

[Dmitro Cashback @DmitroCP]

THEY DROPPED AN ICE STATUE INTO LAVA

20,000 likes. 60 comments.

A woman in a helicopter doorway steadies a carved ice figure, a full human form, then slides it out over an open lava channel. It lands, whitens, and goes up as a column of steam

the choice of object is the entire design, and swapping it for anything else breaks the clip:

→ a plain block of ice in the same shot is physics. a carved figure is a body, and the eye refuses to unsee that

→ the melt therefore reads as something happening to someone rather than to a material

→ the face is the last part to go, which is where the camera holds, and that ordering is deliberate

→ the lava channel is filmed from directly above, so the figure falls into a line rather than onto a surface. it looks swallowed

→ and nothing explains why. no caption on screen, no setup, no reaction shot

this is the difference between a physics demo and an image, and it is where generated video is currently sorting itself into two tiers

the lower tier shows a material behaving correctly and hopes the behaviour is enough. it usually is not, because correct physics is now the baseline rather than the achievement

the upper tier picks an object the viewer cannot look at neutrally, then lets ordinary physics happen to it. same render cost, entirely different result, and the decision is made before anything is generated

worth adding what the caption is doing, because it is a separate trick. the text under this post has nothing to do with the video. it is lifted wholesale from an unrelated cartoon account, in Japanese, to borrow search terms

if you want to test how much the object choice carries, run the same motion on two subjects. image-to-video off one still is the cheapest way to do it. @Picsart runs it from a phone

ice melting is physics. a face melting is something you have to look away from

[Valentin @martynov014]

THEY DROPPED A GIANT MENTOS INTO COCA COLA

22,000 likes. 88 comments

A man in a helicopter doorway wrestles a white disc the size of a table, a Mentos packet scaled up beside it, and pushes it out over a Coca-Cola can built like a grain silo. The foam comes back up through the door and covers him

this is a schoolyard experiment rebuilt at industrial scale, and the scale is doing all the work:

→ everyone in the audience already knows the reaction. no setup needed, no explanation, and the anticipation starts from the first frame

→ both objects are branded, and both brands are things people hold in their hands weekly. that familiarity is what makes the size legible

→ the payoff is physical and instant. foam rising, then the operator getting hit by it, which converts a chemistry demo into slapstick

→ the helicopter framing is reused from the whole lava-drop genre, but the reaction here is the point rather than the drop

→ and it runs thirteen seconds, ending on the soaking rather than lingering after it

what makes this format durable is that the viewer supplies the physics themselves. a generated reaction only has to be roughly right, because everyone is watching to confirm a prediction they already made, not to evaluate a simulation

that is the cheapest possible position for generated video to occupy. no one is checking the render, they are checking whether it matches the version in their head

and the brands are load-bearing again. this shot does not work with a generic mint and a generic can, because the size cue and the shared memory both come from the packaging

which puts a lot of trademarks into a genre nobody licensed them for, purely because recognisability is what makes scale readable

if you want to test how much a known object carries a shot, image-to-video off one still is the cheapest experiment. @Picsart runs it from a phone

the chemistry is real. the only fake part is how much of it there is

[Bob Alice @oscilloo]

COP PULLING OVER A CAR THAT DOESNT EXIST

804,000 likes. 3,737 comments

Highway, ocean on the right, a motorcycle officer gesturing at a matte black wedge of a car with exposed exhausts. Filmed from a passing vehicle

the shot is built entirely out of things that are cheap to render:

→ the officer is seen from behind and at distance. no face, no badge detail, no expression to sustain

→ the car has no interior, no visible driver and no glass to reflect a world correctly

→ both vehicles move at similar speed, so the background blurs uniformly and hides almost everything

→ the framing is a passenger window shot, which justifies the shake, the dirt and the partial obstruction

→ and it runs ten seconds, inside the single-pass window of every current model

none of that requires a frontier model, which is the point worth making on this account. no faces, no text, no long continuity, no locked identity across references

Wan 2.2 is Apache 2.0 with a 5B checkpoint that fits in six to eight gigabytes of VRAM. LTX-2 ships open weights with synchronised audio and runs 4K on a 24GB consumer card. Both handle a ten second vehicle shot with a blurred background without difficulty

the closed models earn their price on hard shots: long takes, locked characters, fifty references, dialogue. this is the opposite of all of that

804,000 likes on a clip that a laptop could produce is the actual headline. the ceiling on reach stopped being correlated with the cost of production some time ago, and most people budgeting for this have not noticed

if you want to test an idea before installing anything local, image-to-video off one still is the fastest route. @Picsart does it from a phone

no face, no interior, no detail held longer than a second. eight hundred thousand likes

[Valentin @martynov014]

https://t.co/I1Z0BwhrSi

[Bob Alice @oscilloo]

THREE SISTERS, THREE LABELS, ONE FORMAT

7,307 likes. 201 comments

Three women come down a staircase one after another, each with a small white caption pinned to her as she passes: an age, a relation, nothing else

strip the people out and look at what the video actually is:

→ a static camera, a fixed frame, and subjects walking through it in sequence

→ a text overlay that appears and disappears on a timed cue, anchored to a position rather than tracked to a person

→ no cuts, no transitions, no colour work, no audio design beyond a licensed track

→ a reveal structure that comes entirely from the order of arrival

→ and a payoff that is just the last label landing

that is the entire build. no model touched it, and none was needed, which is worth saying out loud on an account that spends most of its time on generation

the templates that dominate short form are still overwhelmingly text layers on ordinary footage. timed captions, a countdown, a label pinned to a moving subject. the technology involved is forty years old and available in every free editor on earth

ffmpeg does timed overlays from a subtitle file in one command. Kdenlive and Shotcut do the same with a timeline and cost nothing. None of it needs a GPU, a credit balance or an account

which is the useful reminder. the arms race in generated video is real, and it is also a distraction from the fact that most formats winning right now are won at the text layer, where the tooling has been free and open for decades

worth knowing which of your problems actually needs a model before you go pay for one

if you do need motion on a still, image-to-video is the cheap way to test it first. @Picsart runs it from a phone

three captions and a staircase. no model in sight

[Valentin @martynov014]

https://t.co/I1Z0BwhrSi

[Bob Alice @oscilloo]

ROBOT IN SCRUBS RUNNING THROUGH A PLAZA

22,000 likes. 881 comments.

Humanoid machine in a hospital gown, sprinting past a fountain and some patio chairs, filmed from a distance on what looks like a phone

what makes this work is the shooting style, and it is the cheapest style there is:

→ the camera is far away and unsteady. distance hides the detail a model would have to get right up close

→ the subject is partly occluded the whole time by planters, bollards and chairs. every obstruction is a frame the model does not have to render

→ the scrubs solve clothing. loose fabric over a rigid body hides the joint problem that gives most humanoid renders away

→ there is no face to track. a blank head is the single biggest cost saving available in this genre

→ and the framing is amateur on purpose. bad footage reads as real footage

that is an entire aesthetic built around the model's weaknesses, and it is why the surveillance and bystander look dominates this category

worth saying plainly on this account: none of that needs a frontier model. the shot has no face, no text, no lip sync, no long continuity, and it runs fifteen seconds. Wan 2.2 is Apache 2.0 and its 5B checkpoint fits in six to eight gigabytes of VRAM, which is a laptop GPU

the closed models earn their price on the hard shots, long takes with locked identity across fifty references. this is the opposite of that, and paying per second for it is a habit rather than a requirement

the useful discipline is to sort your shots before you pick your tool. most feed content sits in the cheap tier, and the cheap tier has been running on consumer hardware for over a year

if you want to test an idea before you set up a local pipeline, image-to-video off one still is the fastest way in. @Picsart does it from a phone

shaky, distant and obstructed. every one of those is a cost saving

[SOFIA @SofiaTachExpert]

https://t.co/svQ0jDFIH3

[Bob Alice @oscilloo]

SHE RODE A DRAGON UP THROUGH THE CLOUDS

739,000 likes. 2,288 comments

First person, hands on the reins, a hundred metres of mossy war beast lifting off the ground and climbing into cloud. The caption names the model: Seedance 2.5

thirty seconds is the number to look at, not the dragon:

→ Seedance 2.5 does single-pass clips up to thirty seconds, and this one sits exactly at the ceiling

→ it takes up to fifty references, images, video and audio together, which is how the scales, moss and harness stay identical from the ground to the cloud layer

→ first person removes the face problem completely. two hands and a set of reins is the entire human presence

→ cloud is the friendliest possible background. it has no geometry to keep consistent and it hides the horizon

→ and the climb is one continuous move, so there is no cut anywhere to betray it

what that adds up to is a shot that would have needed a studio two years ago, made by one account, at the top end of what a closed model currently offers

the open side is close enough behind to matter. Wan 2.2 is Apache 2.0 and its 5B checkpoint runs in six to eight gigabytes of VRAM, which is a laptop. LTX-2 puts native 4K with synchronised audio on a 24GB consumer card, weights on disk, no meter running

the honest split is this. thirty unbroken seconds with fifty locked references is still where the paid models win. a ten second shot with two references is not, and most of what fills a feed is the second thing

if you want to find out which half your idea lives in, image-to-video off one still is the cheapest test. @Picsart does it from a phone

a hundred metres of dragon, thirty seconds, one prompt window

[Dmitro Cashback @DmitroCP]

HE DUMPED CHICKEN WINGS INTO A VOLCANO

13,000 likes. 103 comments

A man in a monkey mask sits in an open helicopter door over an active lava field and tips a crate of raw wings out onto the rock

thirteen thousand likes is a modest number, and that is what makes this one worth reading:

→ the format is already saturated. helicopter, lava, food, GoPro angle. this is the fourth variation of it I have logged this month

→ the mask is the tell. when the premise stops carrying a clip, creators add a character to it, and a mask is the cheapest character available

→ the lava is good. glow through smoke, heat haze bending the rock edge, the right dull orange rather than fire orange

→ the wings behave correctly on impact, which a year ago would have been the whole achievement

→ and none of that mattered, because the audience has seen the premise three times already

there is a lesson in the gap between quality and reach here. the render is better than the clips that did twenty times these numbers in August. it arrived late to its own format

formats decay, and they decay fast now, because the thing that used to gate imitation was production cost and that gate is gone. a premise that lands in week one is furniture by week four

so the useful skill stopped being execution. it is timing, and specifically knowing when a format is still ahead of its audience

if you want to test a premise while it is still early, image-to-video off one still is the fastest way to find out. @Picsart runs it from a phone

the render was not the problem. the calendar was

[Bob Alice @oscilloo]

SHE RODE A DRAGON UP THROUGH THE CLOUDS

739,000 likes. 2,288 comments

First person, hands on the reins, a hundred metres of mossy war beast lifting off the ground and climbing into cloud. The caption names the model: Seedance 2.5

thirty seconds is the number to look at, not the dragon:

→ Seedance 2.5 does single-pass clips up to thirty seconds, and this one sits exactly at the ceiling

→ it takes up to fifty references, images, video and audio together, which is how the scales, moss and harness stay identical from the ground to the cloud layer

→ first person removes the face problem completely. two hands and a set of reins is the entire human presence

→ cloud is the friendliest possible background. it has no geometry to keep consistent and it hides the horizon

→ and the climb is one continuous move, so there is no cut anywhere to betray it

what that adds up to is a shot that would have needed a studio two years ago, made by one account, at the top end of what a closed model currently offers

the open side is close enough behind to matter. Wan 2.2 is Apache 2.0 and its 5B checkpoint runs in six to eight gigabytes of VRAM, which is a laptop. LTX-2 puts native 4K with synchronised audio on a 24GB consumer card, weights on disk, no meter running

the honest split is this. thirty unbroken seconds with fifty locked references is still where the paid models win. a ten second shot with two references is not, and most of what fills a feed is the second thing

if you want to find out which half your idea lives in, image-to-video off one still is the cheapest test. @Picsart does it from a phone

a hundred metres of dragon, thirty seconds, one prompt window

[Valentin @martynov014]

https://t.co/I1Z0BwhrSi

[Bob Alice @oscilloo]

T-REX FOUGHT A MEGALODON IN SURF

16,000 likes. 177 comments

Two animals that never met, on a coastline that never existed, in a fight nobody can fact-check

and that last part is the whole reason this genre suits local models so well:

→ no ground truth. nobody knows how a tyrannosaur actually moved, so there is no reference anyone can hold the render against

→ no faces. skin, scale and muscle are forgiving, and the uncanny valley has no entrance here

→ no text, no hands, no lip sync. the three classic failure modes are simply absent from the brief

→ water hides the joins. spray, foam and turbulence cover exactly the transitions that would otherwise look glued

→ and the audience arrives pre-sold. every person watching already ran this matchup in their head as a child

put those together and you get the cheapest possible showcase for a model running on your own hardware. no identity to keep consistent, no physics anyone can check, no dialogue to sync

which is worth stating plainly, because the open-versus-closed argument usually gets fought on the hardest shots. it should be fought on the ordinary ones. a clip like this has nothing in it that a local 5B checkpoint cannot do, and the gap that justifies a per-second meter simply does not appear in frame

the honest limit is that 16,000 likes is a modest number for a monster fight, and that tells you the other half of the story. the technique is free now, so the scarce thing is a reason for the matchup to matter

if you want to see the floor for yourself before installing anything, image-to-video off one still is the fastest route in. @Picsart does it from a phone

nobody can prove it wrong, which is exactly why it was cheap to make

[Bob Alice @oscilloo]

T-REX FOUGHT A MEGALODON IN SURF

16,000 likes. 177 comments

Two animals that never met, on a coastline that never existed, in a fight nobody can fact-check

and that last part is the whole reason this genre suits local models so well:

→ no ground truth. nobody knows how a tyrannosaur actually moved, so there is no reference anyone can hold the render against

→ no faces. skin, scale and muscle are forgiving, and the uncanny valley has no entrance here

→ no text, no hands, no lip sync. the three classic failure modes are simply absent from the brief

→ water hides the joins. spray, foam and turbulence cover exactly the transitions that would otherwise look glued

→ and the audience arrives pre-sold. every person watching already ran this matchup in their head as a child

put those together and you get the cheapest possible showcase for a model running on your own hardware. no identity to keep consistent, no physics anyone can check, no dialogue to sync

which is worth stating plainly, because the open-versus-closed argument usually gets fought on the hardest shots. it should be fought on the ordinary ones. a clip like this has nothing in it that a local 5B checkpoint cannot do, and the gap that justifies a per-second meter simply does not appear in frame

the honest limit is that 16,000 likes is a modest number for a monster fight, and that tells you the other half of the story. the technique is free now, so the scarce thing is a reason for the matchup to matter

if you want to see the floor for yourself before installing anything, image-to-video off one still is the fastest route in. @Picsart does it from a phone

nobody can prove it wrong, which is exactly why it was cheap to make

[Lunar @LunarResearcher]

https://t.co/nN6fldtV4f

[Bob Alice @oscilloo]

@ViajandconFlopi looks scary

[Bob Alice @oscilloo]

@martynov014 @dianycorazon @neymarjr i think it's fake

[Bob Alice @oscilloo]

A BORDER COLLIE TOOK THE WHEEL

219,000 likes. 4,232 comments

Generated, and specifically the kind of thing that no longer needs a datacentre behind it

what this clip demands of a model is a short list, and every item is now cheap:

→ two animals, no humans. fur is forgiving in a way skin is not, and nobody has a precise mental model of a dog's shoulder geometry

→ a car interior, which is a box with known proportions and fixed lighting. one of the most over-represented environments in any training set

→ almost no motion. paws resting, heads turning, scenery blurring past the window

→ no text, no hands, no lip sync. the three classic failure modes are simply absent

→ and ten seconds, which is inside the single-pass window of every current model

which is why this is worth pointing at from the open-source side. a clip like this does not require an API key any more

Wan 2.2 is Apache 2.0, shipped July last year, and was the first video diffusion model with a mixture-of-experts architecture. The 5B checkpoint runs in 6 to 8GB of VRAM - that is a laptop GPU, not a rented H100. Even 14B GGUF builds squeeze into the same range with aggressive offloading

ComfyUI has native templates for it: text-to-video, image-to-video, first-frame-to-last-frame. Nothing to reverse-engineer, nothing to pay per second

the honest framing is that the closed models still win on hard shots. this is not a hard shot. most of what fills a feed is not a hard shot, and that entire tier of work has quietly moved onto hardware people already own

if you want to test an idea before installing anything, image-to-video off one still is the fastest way in. @Picsart does it from a phone

two dogs, one car, eight gigabytes of video memory

[Lunar @LunarResearcher]

https://t.co/nN6fldtV4f

[Bob Alice @oscilloo]

8 GTA CHARACTERS WALKED A RED CARPET AS REAL PEOPLE, ONE AT A TIME, EACH STOPPING FOR THE PHOTOGRAPHERS UNDER THEIR OWN NAME CARD

123,000 likes. 1,354 comments

Michael. Niko Bellic. Franklin. Trevor. Then all of them in one group shot at the end, which is the part that should interest you

because the group shot is the hard problem, and everything before it is the setup:

→ eight separate identities, each one recognisable from a source that is a drawing, not a photograph. the model had to invent a face that reads as that character without ever having seen it

→ every character appears twice - alone, then in the group - and survives the jump. holding one identity across two unrelated shots is the thing that used to collapse

→ each one is matched to their own artwork on the wall behind them. the render has to agree with a reference standing next to it in frame

→ the lighting changes between segments and the faces do not. press-wall flash is brutal, and it is where generated faces usually go plastic

→ and nobody moves much. a red carpet is a format where standing still and turning slightly is the entire expected behaviour

that last one is the craft. the creator picked a real-world ritual whose whole choreography is "stand there and let people photograph you," which happens to be the exact motion budget a video model can afford

the identity problem underneath it is solved, and solved in the open. the working rule people have settled on is: a reference adapter for speed, a LoRA for scale, ControlNet for structure. IP-Adapter behaves like a one-image LoRA - hand it a single clean face and it carries that identity into a generation, no training run required

there is an even cheaper trick going around: instead of fighting for consistency across separate generations, produce one continuous clip that orbits the character 360 degrees. every frame comes from the same generation, so consistency is free, and you harvest your reference sheet out of it afterwards. runs locally, costs nothing per call

which is the real headline. character consistency was the last thing that genuinely required a budget, and it is now a workflow file somebody published for free

if you want to test the idea before building the graph, image-to-video off one still is the shortest way in: one image, one line about the motion. @Picsart does it from a phone

eight faces that have never existed, holding still long enough to be photographed

[Bob Alice @oscilloo]

8 GTA CHARACTERS WALKED A RED CARPET AS REAL PEOPLE, ONE AT A TIME, EACH STOPPING FOR THE PHOTOGRAPHERS UNDER THEIR OWN NAME CARD

123,000 likes. 1,354 comments

Michael. Niko Bellic. Franklin. Trevor. Then all of them in one group shot at the end, which is the part that should interest you

because the group shot is the hard problem, and everything before it is the setup:

→ eight separate identities, each one recognisable from a source that is a drawing, not a photograph. the model had to invent a face that reads as that character without ever having seen it

→ every character appears twice - alone, then in the group - and survives the jump. holding one identity across two unrelated shots is the thing that used to collapse

→ each one is matched to their own artwork on the wall behind them. the render has to agree with a reference standing next to it in frame

→ the lighting changes between segments and the faces do not. press-wall flash is brutal, and it is where generated faces usually go plastic

→ and nobody moves much. a red carpet is a format where standing still and turning slightly is the entire expected behaviour

that last one is the craft. the creator picked a real-world ritual whose whole choreography is "stand there and let people photograph you," which happens to be the exact motion budget a video model can afford

the identity problem underneath it is solved, and solved in the open. the working rule people have settled on is: a reference adapter for speed, a LoRA for scale, ControlNet for structure. IP-Adapter behaves like a one-image LoRA - hand it a single clean face and it carries that identity into a generation, no training run required

there is an even cheaper trick going around: instead of fighting for consistency across separate generations, produce one continuous clip that orbits the character 360 degrees. every frame comes from the same generation, so consistency is free, and you harvest your reference sheet out of it afterwards. runs locally, costs nothing per call

which is the real headline. character consistency was the last thing that genuinely required a budget, and it is now a workflow file somebody published for free

if you want to test the idea before building the graph, image-to-video off one still is the shortest way in: one image, one line about the motion. @Picsart does it from a phone

eight faces that have never existed, holding still long enough to be photographed

[Lunar @LunarResearcher]

https://t.co/nN6fldtV4f

[Bob Alice @oscilloo]

A WOMAN BALANCED ON A ROLLING BLUE BARREL THE WHOLE LENGTH OF A POOL WHILE A CROWD OF HOLIDAYMAKERS WATCHED FROM THE EDGE

230,000 likes. 1,400 comments

She does not exist. Neither does the barrel, the pool or the crowd. The caption names the model it was made on: Seedance 2.5

which is useful, because it means you can price this clip exactly. the published numbers on that model:

→ it went live on August 7, doing single-pass clips up to thirty seconds

→ up to fifty multimodal references in one generation

→ audio generated alongside the picture rather than dubbed on afterwards

→ pricing from about $0.10 per second at 480p up to $2.08 per second at 4K with audio

→ and it ships C2PA provenance data plus filters that refuse recognisable real faces and copyrighted characters

so fifteen seconds cost somewhere between a coffee and thirty dollars depending on tier, and this account ships the format daily. that is an operating line, metered by the second, denominated in somebody else's currency and governed by somebody else's content policy

which is the entire argument for the open side. weights on your own disk mean the meter never runs, nobody decides after the fact which faces you are allowed to render, and no provenance tag gets attached to work you did yourself

the trade is honest, so it is worth stating plainly: the closed model is better today, and it will charge you per second forever. the open one is behind, and it is yours. only one of those two gaps closes on its own

if you just want to test whether an idea works before picking a side, image-to-video from a phone is the shortest route in - one still, one line about the motion. @Picsart does that

the barrel is fake. the invoice is not

[Lunar @LunarResearcher]

https://t.co/nN6fldtV4f