WHEN THE DRAWINGS ARE MADE, THE CHARACTERS AND SCENES ARE WRITTEN, FOLLOWED BY A LONG LINE OF "SUPER-CLEAR, 8K, FILM SENSE, BEST-QUALITY, SUPER-HIGH DETAILS", AND WHEN THE VIDEO IS MADE, EVERY MOVE OF A PERSON'S HAND, TURN AROUND, STEP FORWARD, AND TURN AROUND IS OPENED, FEARING THAT AN AI WILL NOT BE UNDERSTOOD。
These words cannot simply say “wrong”。
The problem is that models have changed。
Previously, the model itself had limited capacity to understand, and a large part of the reminder function was to help it to supplement its capacity. The default quality of the picture is not enough, and we're constantly drawing words; the video doesn't understand the relationship between actions, and we're tearing down a simple action into steps。
Now the new model is already able to understand more complete natural language and is increasingly judging the person, the environment and the movement. If our hints still devote a great deal of space to what the model can do on its own, not only waste information, but sometimes add additional limits to the model。
So it is not a few so-called “new keywords” that really need to be updated to learn the hints。
What needs to be updated is:
The question of what should be the subject of a hint。
I. The so-called “obsolete phrase” is not the pre-existing term that suddenly became obsolete
LET'S START WITH THE AI MAP。
This may have been a very common hint:
A YOUNG GIRL WAS STANDING BY THE STREETSIDE WITH A SUPER-CLEAR, 8K, FILM SENSE, THE BEST QUALITY, MASTER PHOTOGRAPHY, SUPER-HIGH DETAILS, REAL SKIN QUALITY, AND A FINE PICTURE。
Source from webUI
Why is that writing coming
Because the early model itself has limited picture completion. We only told it that “a girl is standing on the side of the street” and that the last picture that we get is not very detailed enough, has a rough background and is unstable, so you're starting to add a whole range of qualitative improvements to the hint。
If the word “high” is not sufficient, add the word “high”。
IT'S NOT SAFE TO SAY "SUPER-HIGH" AND ADD "8K" AND "BEST PAINTING" AND "MASTER PHOTOGRAPHY"。
What these words did at the time was in fact an auxiliary model to improve the picture completion。
But as the model continues to be updated, this is changing。
Many of the image models are now capable of producing, by default, more complete characters, light, clothing and the environment, even without a large body of painting words. In other words, the question of whether “photograms are sufficiently sophisticated” that has in the past needed to be addressed in a focused manner has not been as prominent as before。
At this point, if we continue to fold in the old way:
8K, SUPER-HIGH, EXTREME DETAIL, BEST QUALITY, FILM LEVEL, TOP-LEVEL PHOTOGRAPHY, PERFECT LIGHT..
This does not necessarily lead to a significant increase。
SOMETIMES, MODELS PUSH THESE REQUIREMENTS TO THE EXTREME, MAKE THE PERSON'S SKIN MORE SMOOTH, LIGHT MORE BEAUTIFUL, PICTURE MORE ORGANIZED, AND THEN HAVE A TYPICAL EFFECT: EVERYTHING LOOKS GOOD, BUT IT'S LIKE AI。
SO THE WORD “ADMISSIBLE” DOES NOT MEAN THAT THE WORD “8K” SHOULD NEVER APPEAR AS OF TODAY。
The real change is:
These words should no longer be at the centre of the narrative。
If you really need some sort of clarity, taste or image style, of course you can keep it. But there is no need to fold up a line of similar graphics in order to appear to be indicative。
Now it's more worth thinking about:
BESIDES TELLING AI, "DO IT BETTER," WHAT ELSE REALLY DETERMINES WHAT THIS PICTURE LOOKS LIKE
The focus of the biographic hint is moving from “to improve the quality” to “to create a sense of truth”
Now go write another photo of itAI ImageIt could be a different idea。
We were describing the results:
Film senses, super-clear, top-level photography, perfect light。
Now we can begin to describe how this picture was created。
For example, a girl is still standing by the street。
The past may read:
A YOUNG GIRL IS STANDING ON THE STREETS OF THE CITY, WITH A HIGH LEVEL OF CLARITY, 8K, FILM SENSE, MASTER PHOTOGRAPHY, AND THE BEST DRAWING QUALITY。
Now we can replace it with:
A young girl was standing by the city street in the evening, waiting for a car, holding a mobile phone and occasionally looking down at the news. Photos are taken at close range using a mobile phone camera to maintain an ordinary, real sense of living record, with natural light on the street on the face and hair, and a normal road, vehicle and urban environment in the background。
The most important change here is not the length of the hint or the addition of more photographic terms。
The type of information has changed。
The first is to tell the model:
"I want a very high-quality picture."
The second writing starts telling the model:
What the characters are doing, what time they are filming, how they are filmed, and what they are supposed to be。
This is a very important direction now for the “real sense word”。
For example, we write “professional studios”, which are understood in a more formal, complete and luminous direction; and “cell phone cameras, ordinary true sense of nature”, which is received by the model。
So really effective information doesn't necessarily sound like a specialty。
Many times it's a very common description:
Cell phone camera shot. Daily life record. Normal indoor light. It's natural at night. The character didn't do it deliberately. Like normal photographs。
By Midjourney
Their role is not simply to “reduce the painting”, but to help the model determine what the final picture should be in the real picture。
What really should be changed is the way of thinking。
If you want a commercial product map, then describe it according to the logic of commercial photography
If you want a picture of a person on the way, describe its environment, time and manner
Cameras, lights and characters should be readjusted if they are to generate official magazine character photographs。
In other words.
NOT ALL AI PICTURES MUST PURSUE THE SAME "REAL SENSE."。
What we really need to tell the model is:
Why is this picture being taken like this。
III. VIDEO TRANSFER: Models begin to complete their own actions
This is not a pictureAI VideoThe change in the hint is more obvious。
Former video models are less able to understand, especially when dealing with the movement of people, and we often encounter a problem: if a normal human being fully understands, the model does not necessarily complete the intermediate process。
for example:
The girl heard her call and turned around。
Generated by the old model C.I. 2.1
That is very clear to people. But there may be many problems with early models。
So in the past we used to have to keep one move open:
The person stops his or her original movement, then he or she moves back, the head starts spinning, the shoulder follows, the upper half continues to rotate, his or her step is repositioned and his or her body gradually turns around..
This approach had its raison d'être in the past。
Because when the model doesn't understand the relationship between actions, we have to tell it for ourselves。
But the situation is beginning to change。
The new video model is increasingly capable of understanding movement, inertia, human relationships and continuity of events, and a normal turn, turn back, walk, stop or reach has been able to automatically complete a significant part of the process。
At this point, if one very simple move is still to be broken down into 78 or even a dozen instructions, another problem may arise。
Because the more you write, the more the model needs to meet at the same time。
The head has its own requirements, the shoulder has its own requirements, the hands are required and there is another time sequence for the leg. When there is a slight conflict between these details, the person that is created is prone to a sense of pause, Qatón or a certain execution。
Eventually it didn't look more natural, but more mechanical。
So now when you write video action, you can start with:
"What exactly does this move do?"
Turn to:
"Why does this happen?"
For example:
Girls turn around。
It can be further written:
The girl was moving forward and suddenly heard her being called behind her, she stopped and turned in the direction of the sound, and then saw the person standing behind。
And then we've told the model three key relationships。
The character was doing something。
Why suddenly change the movement。
What happens after the action。
As to how the head, shoulder, waist and legs should fit in at normal turn, for the current model, it can be done properly。
And that's why the video hints are not as short as possible。
Not to delete the two-hundred-word description of past moves into the word "person turns."。
Instead, the text that was originally used to repeat the control of the course of action was put back to more important content。
IV. THE VIDEO PURPOSE IS REALLY SUPPOSED TO BE CLEAR
When the model is able to complete a part of the action on its own, the work that the creators really have to undertake has not actually decreased。
Just the control focus has changed。
In the past, we may have spent a lot of time regulating how the person's body moves, and it is now more appropriate to determine in advance:
What kind of video is this
Where did the story happen
When did it happen
In what way
What's next for the character
What did you get after this mirror
These are things that you can't decide for yourself。
LIKE YOU TOLD AI:
A girl walked into the station。
The model certainly keeps her going。
But it doesn't know whether you're making a short video of life, a skeptical short film, an ad or a drama film; whether it's morning or night, or whether the camera should be watching by the side, like the passersby, or whether it's always hanging around like a photographer。
These are in the video direction。
It is also a turn, and if it is preceded by “hearing that the radio prompt train is about to enter the station”, the logic of action is one state; if it is preceded by “suddenly hearing someone calling his name behind him”, it creates another state。
So now, one of the most important changes is:
LESS TEXT SAYS AI ALREADY KNOWS HOW TO DO IT, MORE WORDS TO TELL IT WHAT YOU REALLY WANT TO DO。
But it's also not going to go to another extreme。
If it's a normal walk, or a normal turn back, it does make the model complete; but if you want the role to pick up a specified product, two people to exchange an item, the effects must start at a certain point in the arm, or the actions themselves are the most important visual effects of the video, then the processes still need to be clear。
So the new idea is not:
GIVE IT ALL TO AI LATER。
They are:
Controlling the direction that needs to be determined by the creator, and properly liberalizing the natural processes that the model can already understand。
V. Current video alerts that can be organized into four really important parts
What should be written by a full video alert if a large amount of space is no longer spent on physical action
It can be organized in four parts:
Content setting, division of materials, overall tone, mirrors。
IT'S ALSO A STRUCTURE THAT IS VERY MUCH WORTH SETTLING DOWN WHEN IT'S WRITTEN IN A MULTI-MODULAR AI VIDEO。
The first is content-setting。
Tell the model at the beginning of the hint, what kind of video is this? For example:
A 10-second video of the actual person. The story occurred on the summer afternoon of City Business Street, where a girl walked alone after her shopping. The whole video was done by following a mobile phone。
This part addresses not how a single shot moves, but the basic message of the whole video。
The type of video, where the story happens, what time, and what method or device is used to film it, should be determined as far as possible in advance here。
Because this information affects all the footage behind。
The second is the division of material。
Now that many videos have been generated, there is no longer only one reference, and we may be able to provide both figures, scenes, costumes, product charts and even action references。
With more material, the biggest problem is that models need to be judged:
Which one of them is in charge of
So don't just write:
See above for video generation。
They should be described separately。
For example:
Figure 1 refers to the identity of five officials, hairstyles and characters of the female owner; figure 2 refers to the environment, construction and road relations of the commercial street; figure 3 refers only to the clothing of the female owner; figure 4 refers to the appearance of the shopping bag in the hands. The material should not change the structure of the scene, nor should it change the appearance of the person。
copy
The division of materials is clear and essentially to reduce interference between different references。
Part III is the overall tone。
That's what the whole video should always be。
For example:
It's got a good sense of commercial advertising at the film level, a clear sense of nature's early summer air, soft and low contrasts, permeability of the dark, luminous light, colour restraint, and natural integration of green, ivory and warm wood. Real fine skin, fruit-coated fibres, beads, refractions of glass and condensed water tastes, light film particles, soft and high luminous dyes, clean, wet, fresh, with real life temperature, while maintaining the euphemism of high-quality drink advertising。
This part is not the same as a separate “film sense” in the past。
It needs to determine the direction of vision and performance that is common to the whole video, so that the multiple lenses that follow do not create a completely different feeling。
After these three parts have been completed, the final entry into the lens。
Spectroscopes still have to be designed, but don't do all the intermediate moves for the model
At the end of the lens, the creator really needs to control the sequence of events and the lens structure。
For example, a simple character clip can be written as follows:
The girls were sitting alone at the window at the coffee shop, with their left hand on their face and looking outside the window, with coffee and mobile phones on the table. The cell phone suddenly turned on the news, and her attention was drawn, and her head down to take a look at the phone, and when she saw the news, she was relaxed. She then dropped her phone, took a cup of coffee and looked out the window again, when she found her friend waiting to appear across the street. She immediately sat up and was clearly happy, putting down her coffee and picking up the bag on her back and leaving. The first half of the camera kept a close view in front of the person's tilt, slightly retreated and moved when she was about to get up, leaving room for the person to move。
There's still movement and camera changes。
But we did not continue:
The first step is how low the head is, when the right hand reaches the cell phone, how the fingers hold the coffee cup, and which feet are raised to support it。
Because for these normal actions, they are already part of a model that can be built on its own。
The mirror really needs us to be clear:
What the first camera man is doing, what changes her state, what the next act is, and when the camera goes into the next section。
In other words, we used to design:
Every step of the movement。
It's better to design it now:
Why is the whole action going on。
It is also, in fact, one of the most understandable things to understand now。
WHEN AI BECAME MORE CAPABLE, IT WAS NOT THAT WE HAD NOTHING LEFT TO WRITE, BUT RATHER THAT THE HINT BEGAN SLOWLY TO CHANGE FROM “OPERATIONAL INSTRUCTIONS” TO “CREATIVE NOTES”。
So is the map。
We used to tell the model repeatedly:
More clarity, more clarity, more precision。
It's time to start telling it:
What I want is a picture of where it comes from and how it's taken。
So when you're done with the hint, you can ask yourself a very simple question:
DID I WRITE THESE WORDS TO HELP AI UNDERSTAND MY CREATIVE INTENTIONS, OR DID I JUST REPEAT THE REMINDER THAT IT WOULD HAVE DONE WHAT IT WOULD HAVE DONE
In the latter case, it could be appropriately deleted。
If some of the information that determines the direction of the picture is not clear, priority should be given to its completion。
The hints are not the more professional and the less advanced。
What really changes is its writing focus。
It used to be more to compensate for model capabilities。
There is now a need to take advantage of the capabilities already available in the model and to leave limited hint space for information that really determines the end result。
This is where, at this stage, graphics and video alerts are most worth updating。