recent AL COMICIt's really hot. But there are not many complete processes that can really be learned and rewritten。
A LOT OF PEOPLE WANT TO GET INTO THE GAME, BUT THEY'RE STUCK IN THE SAME QUESTION: THEY DON'T KNOW WHAT TO DO WITH THE AI COMIC AND HOW TO RUN FROM ZERO INTO A READY-TO-DOWN PRODUCTION PROCESS。
In the previous period, we had compiled a set of real business comics that had been used only for in-house training. The information itself was valuable, and the process had been done。
It was also easy to share this time — hoping to help those who are looking for directions — and to take this opportunity to meet more friends who are serious about comics, willing to communicate and even cooperate。
This case will be used almost on the marketAI Video CreationMainstream tools. Although there are many Agent tools on the market today that can significantly improve efficiency, in order to build the finest, in most cases it is necessary to follow the traditional "Spectroscopy-Video-Clip" production process, which we have repeatedly referred to as the "Text-Text-View-View-Clip" workflow。
From the point of view of basic skills, we hope that you will have the traditional spectroscopy sense, the ability to map out, the ability to write basic hints and the ability to edit, while building up the audio-visual logic and aesthetic skills associated with traditional video。
In this way, the future, whatever the context of the application, can respond。
i. Scripts and Scripts
When we've finished talking about the question, and we've actually got into the field of business, you'll find a real reality:
And most of the time, we're actually skipping the "creative script." What clients or projecters give us is often already a piece of a novel, a written script, or even a spectroscopy。
SO, AS AN AI VIDEO CREATOR, THE WORK WE'RE GOING TO DO AT THIS STAGE IS NOT ESSENTIALLY “WRITING”, BUT “TRANSLATION” AND “RESTRUCTURING”. WE NEED TO TRANSLATE THE TEXT INTO THE LANGUAGE OF THE CAMERA THAT I CAN UNDERSTAND。
We usually deal with two types of situations in the face of different material from the client。
First scenario: a ready-made spectroscopy script
IF YOUR A IS VERY PROFESSIONAL, IT'S A GOOD THING TO ROLL OVER A VERY WELL-DEVELOPED SPECTROSCOPY THAT COVERS THE SCENERY, THE DRAWING, THE IMAGE, ETC. BUT REMEMBER, DON'T JUST OPEN THE AI TOOL TO START THE RUN。
The first thing after getting the lens is communication. We don't have to re-write the script, but we have to communicate to get to know each other. You need to understand the story of the whole film and the worldview。
Is Sabbonk's neon feeling? Or is it the East? It's only through this that you'll be able to grasp the character and art style。
It's important that you come out of the picture, or even if it's right, it tastes wrong。
Despite the demand for the premonitions, the spectroscopy at hand contains five scenes, each of which is a short play, which is less relevant and involves 10 different role images. There is no doubt that this places high demands on us in terms of overall rhythm control, graphic presentation and the level of final editing。
So first we talked to the writer about the world view and the set. You can't see it in the scripts and the spectroscopy. That is why the writers of the traditional film industry read the original novels first when they are translated into scripts。
AND FINALLY, YOU NEED TO DO A PREVIEW OF THESE MIRRORS IN YOUR HEAD, LIKE A DIRECTOR LOOKING AT THE SCRIPT. IF CERTAIN IMAGES ARE FOUND TO BE DIFFICULT TO ACHIEVE UNDER THE EXISTING TECHNICAL CONDITIONS OF AI (E.G. EXTREMELY COMPLEX MULTI-PERSON INTERACTIVE LONG LENSES), THEY WILL HAVE TO BE PRODUCED BEFORE THEY ARE PRODUCED, WITH ALTERNATIVE SOLUTIONS。
Second scenario: the script
This is the more common state. All we have to do is break it down into a standard spectroscopy form, following audio-visual logic。
The header of this spectroscopy table usually includes the spectroscopy number, the view, the camera movement, the image content, the sound/word, the length of time。
WE CAN, OF COURSE, USE LARGE LANGUAGE MODELS TO HELP US COMPLETE THE FIRST VERSION OF THE CONVERSION FROM TEXT TO FORM. YOU CAN USE THE FOLLOWING HINT TO GET AI TO FINISH YOUR ROUGH CUT:
🎨
[Phrasing template: text into the first mirror]
Rowe: Professional short film spectroscoper
Task: Read what I've been providing, and break it down into a detailed spectroscopy form for the AI video。
Reviews:
- Translating logic: The translation of literary writing into visualized lens language and the prohibition of unphotogenic descriptions such as "the mind" must be translated into action or micro-expression。
- Format requirements: Output is the Markdown table, which contains columns: mirror numbers, spectrometry (distant/full/medium/near/special), lens motion (push/ pull/swipe/swipe/tweak/follow/fix), image content (detailed description of the subject, environment, photocopy), sound/line, estimated duration (sec)。
Input:
ONCE YOU GET THE FIRST VERSION OF THE TABLE GIVEN BY AI, IT CAN NEVER BE USED DIRECTLY。
AI PRODUCES MIRRORS THAT ARE OFTEN ONLY LOGICAL BUT LACK A “FILM SENSE”. IT'S TIME TO CALL YOUR AUDIO-VISUAL LANGUAGE. YOU'VE GOT TO KEEP PROJECTING THESE IMAGES IN YOUR HEAD, SEE IF THEY'RE SMOOTH OR SLOW。
In order to address this problem, we need a “replicate” of the lens. In response to the fast-paced nature of the short play, I have refined the lens assembly principles (e.g., kinetics, 30 degree rules, matching principles, etc.) that I learned earlier into a set of structured instructions that big models can understand。
YOU CAN SEND THE FIRST VERSION TO AI AGAIN, AND IF YOU NEED A HINT THAT CONTAINS THE ABOVE-MENTIONED PRINCIPLES, YOU CAN GET IT THROUGH THE END-OF-TEXT SCAN。
If you're just a beginner, the image is not as strong as it is, you can optimise with the above-mentioned hint. However, with the increase in the number of people reading and making films, it is important to consciously pre-enumerate and make manual changes。
WHEN IT'S ALL READY, YOU GET AN AI DECOMPOSED AND OPTIMIZED SPECTROSCOPY SCRIPT, AND WE CAN GO TO THE NEXT STEP。
BUT IT IS IMPORTANT TO REMEMBER THAT THE SPECTROSCOPY IS BY NO MEANS STATIC, AND THAT IT CHANGES THROUGH THE ENTIRE PROCESS OF YOUR PRODUCTION. WHEN YOU CREATE A PICTURE OR VIDEO IN THE BACK, YOU MAY FIND THAT ONE OF THE SHOTS THAT WAS DESIGNED WAS NOT WORKING WELL, OR THAT AI RAN OUT OF A SURPRISE IMAGE, AND YOU HAVE TO COME BACK AT ANY TIME TO ADJUST THE LENS。
If you're a man with the whole process, then you're the director, and you're ready to change and communicate。
II. Role design
If the spectroscopy script is almost established, the next role, scene design will begin. The main thing is role design。
IF YOU'VE OPTIMIZED THE SPECTROSCOPY SCRIPT IN THE AI CHAT WINDOW, YOU CAN JUST SAY TO IT:
Zenium
PLEASE STUDY MY SPECTROSCOPY CAREFULLY, AND MAKE AN AI DRAWING HINT FOR THE CHARACTER I'M DESIGNING THE CHARACTERS THAT APPEAR IN IT. BACKGROUND IS PURE WHITE. YOU HAVE TO INCLUDE A PERSON'S APPEARANCE, CLOTHING, HAIR, ETC. ONE EDITION EACH IN CHINESE AND ENGLISH. PAINTING IS 2D ANIMATION STYLE。
Note: If the full spectroscopy script is already available, then the spectroscopy script will be uploaded directly to tools such as Gemini, and AI will be requested to generate role image tips according to the hints above。
And after that, we had to create six characters, and 10 images。
Ji Fa Fai: The image of an ancient alien woman in a cloak
Older: Images of middle-aged men of the ancient Gentiles in cloaks
Muyun Dynasty: The image of the ancient young generals: The image of the ancient young generals, the broken armor, the mess of their hair, the blood of their blood
Silent Moon: ancient women, red bridegrooms, cape, blood stains
Xiao Lan Moon (closing): ancient women, pink clothes, pale faces
Jade kite: a warm, ancient image of a man
Jade kite (blood) : The warm old man with blood
Sons of a man: the image of an ancient man
Sons of a Dragon: The image of the ancient young emperor
These are human beings and must be communicated well in advance。
For example, the first two characters, Qi Lan and the old:
SHIFUMI AND THE OLD WILL ONLY APPEAR IN THE FIRST ACT, AS A MATCH WITH HEAVEN, AS A GAME OF CHESS, AND NOT AS TRADITIONAL HAN COSTUMES AND FACES, WITH SOME ALIEN COLOURS. SO WHEN WE TALK TO AI, WE'RE GOING TO TELL AI ABOUT THESE PEOPLE。
SO AI GAVE ME A HINT:
Kihua-chan
2D ANIMATION STYLE, AN ANCIENT ALIEN WOMAN, FULL OF IMAGES. SHE WAS ABOUT 20 YEARS OLD, TALL NOSTRILS, CLEAN FACES, WHITE SKINS, DEEP AND FIRM EYES. THE LONG HAIR OF A DARK HEAD IS NATURALLY FALLING, AND PART OF IT IS BOUND BY FOREIGN-STYLE HAIRS, WEARING DISTINCTIVE LITTLE SILVER. HE WEARS A DARK CLOAK, HE HAS A HEAVY BODY, HE HAS A SIMPLE EMBROIDERY ON THE EDGE, AND HE WEARS AN ALIEN-STYLE TAN DRESS. PURE WHITE BACKGROUND。
Old man
🎨
2D COMIC-STYLE. FULL-BODY, FRONT-LOOKING, AN ANCIENT FOREIGN MAN AROUND 40 YEARS OF AGE. HIS FACE IS STURDY, HIS EYES ARE WISE AND DEEP, AND HIS NOSE IS HIGH. IN A TRIBAL-STYLE DARK ROBE WITH A THICK DARK CLOAK ON THE EDGE OF WHICH THERE IS A SIMPLE TOTEM EMBROIDERY. HE HAD IN HIS HAND A LAVISH WAND WITH STONES OR BONES ON ITS TOP, WHICH SEEMED MYSTERIOUS AND POWERFUL. PURE WHITE BACKGROUND
Our previous articles have shared how you can generate a one-time, positive + three-viewer text template。
Zenium
Generate a three-dimensional view of the body and a facial close-up. (The top left is full of one third of the super-large facial features, 2/3 of the right is positive view, side view, back view) + "Personal description"
AND WE CAN ACTUALLY CREATE THE SIDE, BACK, NEAR VIEW IN TURN, AFTER THE FACE OF THE MAN. FOR EXAMPLE, WHEN WE'VE PRODUCED A FULL-FACE PICTURE OF A SATISFACTORY CHARACTER, IT'S UPLOADED DIRECTLY INTO THE DREAM AI, AND IN TURN, IT'S ENOUGH TO CREATE A SIDE/BACK/NEARVIEW。
Why is there a series of positive/side/back maps here? Because we use different tools, we play different ways。
One of the characters, Aipai and Agent, is better suited to one picture, and of course we put the different perspectives in a picture as far as possible。
In the case of Vidu, the Vidu, the Clin, the AI video sites, they are often able to upload multiple images as the same role, so we can upload character maps from different perspectives。
The rest of the players are also using this operation to generate images of different perspectives, and can still be a three-way view, depending on the platform you use。
IT'S WORTH MENTIONING THAT YOU DON'T THINK THAT A.I. GAVE YOU CHARACTER TIPS THAT COULD BE PUT IN PLACE ONCE. DON'T SAVE THIS TIME. PRE-DEFINING THE ROLE IS NECESSARY. WHILE AESTHETICALLY, AS FAR AS POSSIBLE, THE SCRIPT IS DESIGNED TO MEET THE NEEDS OF THE A。
After communicating with the writer, we optimize the hints for each role, a process that is artificially and continuously attempted, for example, in the form of a hairdresser and the quality of the whole person。
In hairstyles, men of ancient times usually have their hairs tied after the 20s, and the role of a general, like Muyun, must be attached to the hair, not to be too gorgeous, and therefore not to have hairs。
The kite is made up of famous doctors, whose overall accommodation is not good, and also has a sense of a hermit, whose hair can be a hair + hair。
There were two scenes of the day of the wedding, when the family was set up and turned into a prisoner, so the second scene was set up with her wearing a broken red wedding dress, with her head covered and her ring dropped。
As a member of the royal family, the son is dressed as a nobleman. He can have the jade, and he wears gold. As the man behind the scenes, he can look a little dark in his eyes。
The fourth act of the Lunar Moon was to shape her presence, with no pink words, a sick face and a very light pink daily dress。
The son-in-law had been crowned in the last act, at which time he should have worn a dragon robe and a coronation/colour。
Communication was in place for the preceding period and the role was finalized relatively smoothly, with the parties adopting it and producing a spectroscopy for the next step。
III. Video presentation
The figure is closely linked to the video, which is actually used for video. In this section, we will dismantle some of the key techniques of this film in terms of drawings and videos。
In the past, it was common practice for a pure video workflow to generate pictures with a separate hint for each spectroscopy, followed by one by one. Today, however, models and tools have made significant progress compared to a year or even a half ago. A number of processes have been optimized and production difficulties and thresholds have been significantly reduced。
The film was produced in October, when models such as Nano Banana Pro, Clint O1, Dream 4.5, Seach 2.3 and Wan 2.6 were not published. Compared to the model of 11 – December, it was much more difficult at that time to figure out. The tools on which we rely more are dreams, Vidu, Coven, conch and PS。
For the time being, the model used by the Compatible Aesthetics and applicable to images of this type is the Seedream 4.0, or the dream 4.0 model in Dream AI, given the demand for a 2D animation style. It has a much better understanding of Chinese antiques and dress tips than traditional drawing models such as Midjourney v7. Of course, you might also like to try open-source mapping models, such as liblib, stars, runninghub, and so on, without demonstrating。
NORMALLY, DREAMAI HAS MET OUR NEEDS。
Turning to the chart, we first have to get a clear picture of what is necessary。
Why think about it? This is because video models are becoming more powerful today, such as Vidu q2's reference video, Sora 2's role reference, and recently released video models such as Vidu O1, Wan 2.6, are constantly being upgraded. Many of the spectroscopes are now available in the form of the "Facult video + video frame optimization" and can even directly use role reference to generate video material. As a result, much of the video material does not have to come out of the picture before it is generated。
In the past, we worked hard to figure out, not only to accurately control the role, scene and style of the video. As the tools and tools continue to be updated, we naturally have to continuously optimize our own work processes。
One, straight out
Return to the question just raised. For those images that do not contain humans, more abstract ones, or visions, or landscapes, we can all consider the use of the traditional method of direct generation of hints to pre-empt the drawing。
These cameras, which often do not contain too complex a person's move, position, are essentially intended to display an environment in which the aesthetic requirements for the image as a whole are higher than for video motion. Image models are certainly better than video models. It is also available in subsequent reference videos。
1) AI AUXILIARY GENERATING SPECTRUM PHRASING
AND WE'RE GOING TO CONTINUE IN THAT CHAT WINDOW THAT GENERATES THE SPECTROSCOPY SCRIPT, AND WE'RE GOING TO SORT OUT THE NEEDS SO THAT AI CAN HELP US THINK ABOUT ALL THE SPECS. NOW, LET'S NOT MAKE A DISTINCTION BETWEEN WHAT'S DIRECTLY GENERATED BY THE HINTS AND WHAT'S NOT. LET'S LET'S GET AI TO LIST THEM ALL。
AI SHOULD GIVE A FULL AND CLEAR MESSAGE. YOU CAN BOTH COMMUNICATE WITH AI ACCORDING TO THE FOLLOWING TEMPLATE FOR THE MESSAGE, AND OPTIMIZE THE FOLLOWING TEMPLATE ACCORDING TO YOUR OWN DETAILED NEEDS:
💡
Please recall the final version of the spectroscopy table in front. Give me an ai-painting words for each spectrograph against the contents of the table and include them in the table. Note: 1, the hints should include scenes, scenes, lens angles, image content, overall color, etc. 2. The contents of the spectroscopy cannot be simply collated, but you need to extend them in a logical manner. 3. Each spectrometer, where a hint cannot be used only by name to designate a character, and contains a specific character, and by a specific spectroph to describe his appearance, (the spectra of the same role in each spectrometer is to be consistent). Four, each spectroscopy will be independently generated by the Ai painting model, and AI does not understand the context of the spectroscopy, and you write a hint to avoid a finger description of the previous specs, such as "or the previous XX", "that woman" and "XX in the previous spectroscopy" and each spect will be clearly described by a hint。
5. The purpose of the diagrams is to pave the way for subsequent video production. So, when you write a hint based on the visual description of the lens, you have to think about what kind of picture you generate, which should be the picture before the action described in the hint. 6. English version of each of the notes. 7 and 2d. Role A:XXX Role B:XXX
Note: Point 6 contains a Chinese and English version of the introduction, which is intended to facilitate the direct application of relevant images if you wish to use them in such tools as Midjourney, Stable Difffusion. The eighth point was to prevent AI from forgetting what had been said, so that some of the main appearance features of the key players were re-listed。
AS YOU CAN SEE, AI HELPED ME TO LIST IN ENGLISH ALL THE ALLUSIONS TO EACH OF THE SCREENSHOTS。
Two: Dream versus Midjourney
For this film, the opening board is a very abstract image. It's definitely the best thing to do with a hint。
The following hint is an optimisation based on the first version of the A.I.P
2D COMIC-STYLE, SUPER REALISTIC. AN EVEN IMAGE. A HEALTHY, WELL-BALANCED MALE HAND-WRITTEN, FLOATING FROM THE VANTAGE POINT ABOVE THE PICTURE, LIES RIGHT ABOVE THE OLD BOX OF ROUND CHESS, WITH A SMALL FINGERTIP, READY TO GO AND PICK UP THE PIECES. THE BOX IS FULL OF DARK AND ROUND BLACK PIECES, AND THERE'S SOMETHING HIDDEN NEXT TO THE BOX THAT IS A BOARD OF UNBRIDLED AND ABSTRACT. THE BOARD CONSISTS OF A DEEP STAR-RIVER REFLECTION, WITH NO VISIBLE PHYSICAL BORDER, AND THE EDGES ARE CONNECTED TO ABSTRACT MOUNTAIN RIVERS, WITH ONLY THEIR VAST SURFACES CLEARLY REFLECTING STARLIGHT. THE VOID UNIVERSE'S BACKGROUND, THE SHALLOW HORIZON, THE WHOLE BLUE AND DARK COLOR, AND THE GLACIAL LIGHT OF THE GALAXY ARE THE MAIN SOURCE OF LIGHT, CREATING AN EXTRAORDINARY, SACRED AND EASTERN AESTHETIC MOMENT。
The comparison is still clear, and the same characters are shown in their characters:
In comparison, Midjourney's style is stronger and the whole style is more artistic. Its understanding of such Chinese wind elements as chessboards, chess, Chinese uniforms, etc. is less than ideal。
AS A RESULT, WE HAVE DECIDED TO USE DREAMAI AS THE MAIN MAPPING TOOL FOR THE PROJECT, FOLLOWING THE PREVIOUS CHARACTER DESIGN AND MULTIPLE ROUNDS OF SUBSEQUENT SPECTROGRAPHS。
3) Role reference + consistency of hints
For some role actions, a relatively simple chart can also be straight out with a hint + role reference。
DREAM 4.0 at the time was already doing a good job in terms of reference role profiles, and now dream 4.5 is doing a better job。
Upload a single-person graph reference for Zihua and the old, and then describe in the hint:
The view, the image, the style are well understood. It's good to be consistent。
(4) Perspectives and Contexts
We can also give priority to the direct generation of pictures with hints for images that do not contain people, or for images like big visions, panoramas。
BECAUSE SUCH CAMERAS USUALLY DO NOT HAVE COMPLEX CHARACTER PERFORMANCES, THEIR CORE TASK IS TO DISPLAY AN ATMOSPHERE AND IMAGE BEAUTY. IN TERMS OF CURRENT AI TECHNOLOGY, SPECIALIZED MAPPING MODELS CONTINUE TO BE USED IN VIDEO-TELEGRAPHING ON THE PRECISION OF THE IMAGES, PHOTO-PHOTOGRAPHIC SENSES AND ARCHITECTURAL AESTHETICS。
Video models tend to sacrifice the details of the context in order to ensure dynamic flow. And we start with a mapping model to generate these big scenes, not only to get a wallpaper-grade high-level map, but also to use it as a “floor map” for the subsequent video, so that the video is clear。
It's a classic big vision. The main subject of the picture is the steep mountains and lightning, and the characters are just two tiny clippings. If video models are generated directly, the texture of the rocks and the light of lightning can easily be lost。
I entered the following hint in my dream:
Look at this picture, the quality of lightning tearing into the night sky, as well as the screeching contours of the starboard in reverse light, are very clear and beautiful, and this is a level where it is very difficult to get out of the current video model。
5) Special drawings
For some of the more specific spectroscopes, the more sophisticated effects can be achieved through the direct generation of the hints。
FOR EXAMPLE, THERE IS A CLOSE-UP CAMERA: FROM A POV PERSPECTIVE, IT FOCUSES ON A MAN WHO LOOKS AT THE DAGGER IN HIS HAND AND REFLECTS THE LIGHT ON HIS FACE. AT THIS POINT, IT IS ONLY NECESSARY TO UPLOAD A ROLE PICTURE AND DESCRIBE THE HINTS IN DETAIL。
Consistency of complex lenses and spectrometry
When simple images are completed, the rest are often complex images that require a strong sense of mirroring, that want to have the effect of a single lens; or that require high consistency in multiple continuums。
IN BOTH CASES, IT IS EXTREMELY INEFFICIENT AND DIFFICULT TO ENSURE CONSISTENCY BY SIMPLY WRITING HINTS TO A BLIND BOX, AND WE NEED TO USE THE AI TOOL TO CONTROL AND DERIVE THE IMAGE MORE PRECISELY。
The first is a long shot or a big dynamic mirror for the end of the lens, at the heart of which lies the clever application of the first frame。
IN THE PAST, WE USED TO USE THIS FUNCTION TO CONTROL IT BY A SIMPLE HINT: TO GENERATE A FRAME A INDEPENDENTLY, TO CREATE A FRAME B INDEPENDENTLY, AND THEN FORCE AI TO CALCULATE THE MIDDLE. BUT IF A AND B'S BACKGROUND IS TOO DIFFERENT OR THE PERSON'S ATTITUDE IS TOO DIFFERENT, AI CALCULATES AN INTERMEDIATE PROCESS THAT TENDS TO COLLAPSE BECAUSE AI CANNOT UNDERSTAND HOW TWO IMAGES THAT ARE NOT LOGICAL CAN TRANSITION。
Here are two common solutions:
1) Redraw the original map
The first is a redrawing of the original map, which applies to the camera position, which is essentially unchanged, but changes in the scenery of the subject。
For example, there was a shot: a blooded dagger fell in the snow and then a pink, white flower grew in the blood。
If the images were to come straight to the surface with simple hints, the texture of the snow and the dagger would never have worked. The resulting video, although it can also be silky gradients with current modelling techniques, changes in context, object and certainly not in line with the situation。
In this case, only the elements that need to be changed should be changed, and for elements that do not need to be changed, such as background, dagger style, consistency must be controlled。
We'll start with a hint to create this image in the middle with a blood dagger, and then the two images will be partially adjusted by dream, bean bag, Nano Banana, or even PS。
AFTER THAT, WE JUST HAVE TO PUT THE THREE AT THE END OF THE PICTURE, AND WE'RE GOING TO HAVE TO CREATE TWO VIDEO PIECES OF THE END FRAME, SO THAT WE CAN PUT TOGETHER A DAGGER, A BLOODBATH, A FLOWER. BECAUSE FIGURES 1 AND 3 ARE MODIFIED DIRECTLY ON THE BASIS OF FIGURE 2, THE SURROUNDING SNOW, LIGHT, IMAGE AND FIGURE 2 ARE PIXEL-SCALED. TO FEED THESE TWO IMAGES TO THE VIDEO MODEL, AI JUST NEEDS TO FOCUS ON THE PROCESS OF CALCULATING THE GROWTH OF THE FLOWERS, AND THE VIDEO IS GENERATED BY A NATURAL SILK SLIDE。
After that, I expect the kite outside the picture to bend over and take the flower, pick it up and elaborate it. At this point, the picture simply ensures that the flower is consistent and that the background is essentially snowy。
So, we can upload the bouquet, the picture of the kite, and the view, and then, the hint says:
Zenium
2D COMIC-STYLE, CLEAN LINE, NEAR VIEW. THE MAN IN THE REFERENCE FIGURE, WITH ONE HAND, GRIPPED THE PLATINUM FLOWER IN FIGURE 1. THE BACKGROUND IS A FIELD SNOWSCAPE AND THE OVERALL COLOR IS CHANGED TO PURE WHITE。
When you get two graphs, you control it through the front frame, you enter a hint, you generate a flower-picking video, and it's silky。
the camera followed a long hand into the picture and took the flowers off. the hand is owned by a young man in a light-green man's robe. set in the frame. 2d animated。
For those scenes where there is a big dynamic mirror, and where there is a significant change in context perceiving relationships, if there is a great difference in perceiving relationships between images, there is a need to ensure that there is a very obvious same element in the first two frames, such as a role, an object, etc. This requires that we properly process the tail map, with tools such as PS, Nano Banana, or dream, to add the same elements to the frame。
For example, the petals switch:
Butterfly turns:
2) Spectrum Derivatives and Perspective Re-engineering
The second technique, which we call spectroscopy derivatives and re-prospectives. It focuses on the challenge of how to keep the picture consistent when there is a change of location (spectroscope)。
FOR EXAMPLE, THE FIRST SHOT WAS A MAN SITTING IN FRONT OF A CHESS BOARD WITH A CHESS PIECE IN HIS HAND; THE NEXT SHOT YOU WANT TO CUT INTO A CLOSE-VIEW CLOSE-UP, FOCUS ON THE HAND OF A CHESS PIECE, OR EVEN SWITCH TO A PIECE OF HIS FIRST-PERSON PERSPECTIVE。
IF THIS LENS IS RECREATED BY SIMPLY WRITING A HINT, THE RANDOM PROBABILITIES OF AI CAN MUTATE THE ATMOSPHERE, THE LIGHT DIRECTION AND EVEN THE WRINKLED DETAILS OF THE PERSON'S CLOTHES, LEADING TO A VERY “JUMPING” WHEN THE TWO IMAGES COME TOGETHER。
The best approach is to “derived” on the basis of the perfect medium map (mother map) that already exists. There are two paths, which are different。
The first is the mainstream “photo-editing method”, which is applied to the lenses of small changes in the location or simple spatial relationships。
Photo-editing models (e.g. NanoBanana Pro, Ali Millennium Qwen-Edit, Dream Picture 4.5) are already very powerful. If you just want to turn the landscape into a medium view, or to turn the medium view into a close-up, or to move a little bit from the current perspective, to throw the original picture directly into these tools, with simple hint instructions, you can simply change the painting or the focal area, on the basis of maintaining the original image as it is. This method, which is extremely precise and does not lose paint quality, is currently the preferred option for handling simple spectroscopy switching。
However, there is a clear limitation to the image model: the space imagination is inadequate。
When you switch from a very difficult angle, like a shoulder lens, or want to switch the plane from a person's head to a side, a picture model tends to turn the "person" around, but the context is still in the dark. This leads to images that look like people turning around in a paper background, which is very counterproductive。
As can be seen, whether it be a dream or a Nano Banana pro, the picture of this complex background, when I'm going to switch the back, the side, is basically just a person, but the background remains the same。
OR, I'D LIKE TO PRESENT A CLOSE-UP OF A MAN'S POV PERSPECTIVE (THE FIRST PERSON'S PERSPECTIVE) IN THE VIEWER'S HAND, WHICH IS A MUCH LARGER SHIFT IN THE PICTURE THAN A SIMPLE SHOT PUSH。
As you can see, neither the dream nor the Nano Banana pro can come out of a close-up of the first person I want to see in the game。
At this point, video models can be used to recreate the perspective。
THE BOTTOM LOGIC OF VIDEO GENERATION IS BASED ON AN UNDERSTANDING OF PHYSICAL SPACE. WHEN YOU NEED TO CHANGE YOUR POSITION SIGNIFICANTLY, OR TO SWITCH YOUR COMPLEX PERSPECTIVE (E.G. THIRD PERSON TO FIRST PERSON TO POV), YOU CAN USE THE ORIGINAL MAP THAT YOU HAVE IDENTIFIED AS A BASE MAP, THROW IT TO A VIDEO MODEL, ENTER A MIRROR COMMAND (E.G., "SMART TO HAND-WRITTEN" OR "TO-PERSON PERSPECTIVE") AND LET AI RUN A VIDEO FIRST。
Video models like Dreams, Mirrors, Sora and others have spectroscopy functions, while those models that can be measured in the coven and conch can't be spectroscopyed, but can only change the view and the perspective by delivering mirrors。
BY WAY OF EXAMPLE, AS DREAMAI, AFTER UPLOADING THE MEN'S MID-VIEWS OF THE CHESS PIECES TO THE GRAPHIC VIDEO, WE CAN CHOOSE A VIDEO MODEL OF MORE THAN 3.0 AND THEN ENTER THE NEXT HINT, WITH THE FOLLOWING EFFECT:
AS YOU CAN SEE, THE THIRD SECOND OF THE VIDEO HAS SUCCESSFULLY SWITCHED TO A MAN'S POV VIEW OF CHESS。
While this first video may have a slight distortion of character or environmental details, the video model is based on the physical space of the original map to develop a new perspective, so the images it generates are largely consistent with the previous lens in the environmental atmosphere, the photologic logic and in-depth extension of the background。
How? Imports video material into a editing tool such as a clip, finds the most satisfactory frame, and selects the silent frame above the right。
ALL WE HAVE TO DO IS TAKE THE MOST SATISFACTORY FRAME IN THE RESULTING VIDEO, USE IT AS A BASE MAP, THROW IT BACK IN THE PICTURE TOOL FOR A LITTLE BIT OF DETAIL, OR USE TOOLS LIKE PS TO OPTIMIZE IT, SO WE CAN GET A PICTURE OF THE PERFECT, LOGICAL, AND LOGICAL。
For example, the fact that this cut-off is the right one, but the details in the picture are problematic, such as the fact that there should be petals on the chessboard, the fact that there should be no borders on the board, and the fact that there is a modern-style floor brick in the picture, is a flaw in the picture。
CONTINUE TO USE AI REDRAWS IN TOOLS SUCH AS DREAMS, FIRST TURNING THE BOTTOM OF THE BOARD INTO A TABLE OF STONE AND REMOVING THE FLOOR。
Then you continue to upload to the dream, click on the thumbnail of the picture and set the smart reference area. A brush can be painted where it needs to be modified on the picture, and then enter " a pink petal on the board and the rest will remain unchanged " in the hint。
AFTER THIS PROCESS, WE CAN GET A VERY SMOOTH CONTINUOUS LENS: A CLOSE VIEW -- THE PETALS FALL INTO THE BOARD OF CHESS -- A MEDIUM VIEW -- THE MAN TAKES A PIECE OF CHESS FROM THE BOX, A CLOSE-UP -- A VIEW IN THE EYE
It is not difficult to see that the next few drawings are also captured by video-based models, then optimized, and similar steps are not repeated。
3. Reference student video
THIS IS THE MOST EFFICIENT AND MOST PRODUCTIVE LINK IN THE CURRENT AI VIDEO WORKFLOW. ITS CORE LOGIC IS TO SKIP THE SPECTROGRAPH STEP AND ACT DIRECTLY WITH ROLE ID。
With the introduction of the "subject reference" function in models such as Vidu, we no longer need to draw painful cards for every complex interactive lens. All you need to do is upload the role three view or standard that was previously set to the main library, upload a background map, and give the rest to the hint。
THIS HAS GREAT ADVANTAGES IN DEALING WITH MULTIPLE INTERACTIONS OR DIFFICULT ACTIONS. FOR EXAMPLE, THERE IS A NEED TO EXPRESS “A VIOLENT FIGHT BETWEEN TWO PEOPLE” OR “A DELICATE EMOTIONAL PULL” IN A FILM, WHICH, IF GENERATED BY A DRAWING MODEL, CAN EASILY BECOME BODY-TRAPPED OR DEVIANT. BUT BY REFERENCE TO THE LIVING VIDEO, WE CAN DIRECTLY DESCRIBE THE LOGIC OF THE ACTION, AND THE VIDEO MODEL WILL DRIVE YOUR UPLOADING ROLE ID INTO A NATURAL PERFORMANCE IN THE SCENE。
The bottom left is a screenshot of video material generated using Vidu's reference video function, and the right is a direct picture of the dream 4.5. At first, we tried to produce a direct left effect, and we were hoping to have a left image of a shoulder lens -- Muyun's one-handed neck。
Passer lenses, past and present, are a more difficult perspective to achieve, with a high rate of cards, and even if the pictures are now modified by Nano Banana pro, they often require multiple attempts, and the end result may not be optimal。
It is not difficult to find out that, using the same hint, reference videos directly achieved the effect I wanted, and this step was also avoided. It's not as good as it looks。
Musical design
As this is a real business order demand, original music with copyright should be used as far as possible in music. In AI music tools, Suno is the best。
The use of Suno has been described in detail in previous public articles. If you have a complete spectroscopy script, you can have AI analyze the spectroscopy content and generate consistent Suno style tips for you。
WHEN THE FILM WAS PRODUCED, THE SPECTROSCOPY WAS COMPLETE, SO WE TALKED DIRECTLY TO AI ABOUT THE MUSICAL NEEDS OF THE FILM. AS THE FILM IS DIVIDED INTO FIVE SCREENS, THE LINKS BETWEEN THE SCENES ARE NOT VERY CLOSE AND EACH SCENE HAS ITS OWN UNIQUE EMOTIONAL ATMOSPHERE. SO, BASED ON THE TONE OF EACH SCENE, I WAS ABLE TO COMMUNICATE WITH AI IN ADVANCE, IN COMBINATION WITH A PREVIEW OF MY MIND。
Zenium
remember i sent you the master specs script. i'm going to use the ai to generate the right music. by suno. please write available styles tips for me. pure music. my script is roughly five。
Act I: Set the scene, mood: overall mystery, grandness, destiny, increased tension
Act II: Layout, Emotion/Climate: Cold, Suspicion, Pity, Aura, Grievance
Act three: Conundrum, emotion/temporal: wackness, mind fighting, lumbering from the face, dark flow
Scene 4: End, Emotion/Climate: Grievance, Singing, Grievance Return, Way Forward
Act V: Deciding, Emotion/Climate: Power, Power, Finality, Return to Peace
Note: It would be great if you could see the picture in your head; if you don't, you can write it, just let Suno think for you。
As you can see, Suno has given each of the act's hints, and has given them an analysis so that I can filter and optimize them. Of course, for those who do not normally have music expertise, it is not necessary to do so professionally, so you can just use the hint to "scratch" in Suno, and basically get the right music a few times。
Suno produced these songs, and selected three of them。
This song that was chosen for the first act is a perfect setup. It should, in a sense of hearing, be guided by a sense of quality that, rather than be indiscriminate in the first place, drags the audience into a world known as the sky with a fateful whisper。
The second and third scenes, although they ranged from a dark torture chamber to a seemingly elegant courtyard, corresponded to the atmosphere of the two scenes with a constant sense of suspicion and oppression。
The fourth act is the flashpoint of emotion, with a sense of sadness to live and die, and the song that comes out of it fits. The second half of the song became more emotional and corresponded to the image of the fifth act of the Emperor, who was in the top seat and placed the seal. So this song was chosen in act four and act five, which went straight to the end, running through the latter two scenes, and directly solved the problem of the narrative fault。
Musical choices are not much of a trick, but the key is to understand the emotions that are conveyed in the picture。
ATTEMPTS ARE CONSTANTLY BEING MADE IN THE COURSE OF MANY INTERACTIONS WITH AI. SOMETIMES A SONG IS PRODUCED, WITH ONLY A FEW DOZEN SECONDS OF CLIPS THAT FIT THE PICTURE; SOMETIMES A DIFFERENT PIECE OF A SONG IS USED FOR DIFFERENT IMAGES IN THE FILM. THIS THEREFORE REQUIRES THAT THE CREATOR BE PATIENT, OF FINE TASTE AND CAREFULLY SCREENED。
V. Conclusion
AFTER THIS SET OF PROCESSES, THE MOST INTUITIVE FEELING IS THAT AI ACTUALLY SAVED US A LOT OF WORK, BUT IT STILL CAN'T TAKE OUR PLACE. WHILE THE CURRENT MODEL CAN PRODUCE IMAGES AND VIDEOS QUICKLY, IT IS AT THE HEART OF THE AUTHOR ' S INTUITION OF THE AUDIO-VISUAL RHYTHM THAT IT IS NECESSARY TO COMBINE THESE SCATTERED MATERIALS INTO A LOGICAL AND CONSISTENT STORY。
IN PRACTICE, WE ACTUALLY PLAY THE ROLE OF FILTER AND REVISER. THE CONTENT GENERATED BY AI IS OFTEN RANDOM, WHICH REQUIRES US TO USE THE AESTHETIC AND SPECTROSCOPY CONSCIOUSNESS MENTIONED EARLIER TO DETERMINE WHICH IMAGE WORKS AND WHICH DETAIL NEEDS TO BE REPLAYED。
This death toll in detail is the key to distinguishing between waterline works and fine。
The stronger the tool, the more irreplaceable this professional-based judgement becomes。
So, for friends who have just started trying to make a full-link, I suggest not rely too much on a particular tool, but rather devote more effort to understanding lens language and editing logic。
Even if the AI model evolves again in the future, and there are a variety of one-clicked Agents that can understand even directly complex commands, it is suggested that you do not lose this personal habit。
Technology will be refurbished every few months, but the creative logic that you have accumulated through hands-on is common. As long as this bottom-up logic is in place, and no matter how far the tools evolve in the future, you can do what you really want。