AI Video Tutorial

Why do some of the A.I. characters appear for only a few seconds and we can remember? The reasons are often hidden in several clear features: a slightly exaggerated haircut, carefully dressed as a meeting, accompanied by an unexpected move. The contrast between the person's appearance and behaviour adds a little bit to the ordinary scene. In this section, we're going to design a visible role from zero, and then let him do a natural, rhythmic move. Without Higgsfield members, it is also possible to use available reference material to complete the synthesis of characters in their usual canvass. I know the division of material and the verbs

An ordinary video may also contain stories that are suitable for short films: staff members carry containers into the back of the kitchen, drop them down unexpectedly and then panic and look; a person stands in front of the door, looking at his cell phone, and suddenly turns around, and the dog behind them follows. These are not complex, but they have actions, changes and reasons to keep them alive. The problem is, it's hard to say what we really want to change by writing "to make it more filmy." This section, from material organization, story-breaking, mirror design to production, leads you to vague aesthetic requirements, and turns them into operative hints. Even without the basics of photography, it's..

Now video models can understand more and more natural language, especially in Seedance 2.5, and can produce a fairly real picture of a real person, even if they do not prepare the whole scene in advance, provide a character, then describe space, light and lens by text. This time, I want to take advantage of that capacity. The project as a whole will not produce a large number of reference maps in advance. When actually entering Seedance 2.5, each set of videos uses one of the current person's assets to stabilize the role, outside the person's buildings, space, natural environment, light, video and fine-tuning..

NOW, THE AI VIDEO MODEL HAS BEEN ABLE TO MAKE VERY RICH FACES. THE ROLE FROWNS, SMILES, TEARS, AND COMPLETES VERY SMALL MOVEMENTS OF SIGHT, MOUTHS, HEADS, TURNS. BUT WHEN PEOPLE ACTUALLY START MAKING THEIR EMOTIONAL SCENES, MANY WILL FIND A STRANGE QUESTION: WHEN PEOPLE DO WHAT THEY DO, THE PICTURE THAT EVENTUALLY COMES OUT IS STILL CLEARLY “AI FEELING”. THIS PROBLEM IS EVEN MORE EVIDENT, ESPECIALLY WHEN IT COMES TO STRESS, EMBARRASSMENT, HAPPINESS AND ANGER. PEOPLE START TO LOOK RIGHT AND RIGHT AS SOON AS THEY GET NERVOUS, FROWN THEIR EYES AND SMILE AS SOON AS THEY'RE HAPPY. LOOK AT THESE MOVEMENTS ALONE..

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 IMAGE DEFAULTS ON QUALITY, AND WE CONTINUE TO DRAW WORDS; THE VIDEO DOESN'T UNDERSTAND THE ACTION, AND WE MOVE A SIMPLE MOVE..

IT'S NOT HARD TO MAKE A NICE PICTURE OF AN AI, IT'S A GOOD FIVE-SECOND SHOT, AND IT'S GETTING EASIER. THE REAL DIFFICULTY IS WHETHER THE PICTURE REMAINS STABLE WHEN A PERSON NEEDS A DOZEN SHOTS IN SUCCESSION, WHEN THE STORY GOES FROM INSIDE TO OUTSIDE, WHEN THE PERSON BEGINS TO CHANGE CLOTHES, GET INJURED, SPEAK AND INTERACT WITH OTHER ACTORS. MANY OF THESE PROBLEMS HAVE BEEN ENCOUNTERED, AND THE FIRST REACTION HAS BEEN TO CONTINUE TO MODIFY THE HINTS. WHEN PEOPLE CHANGE THEIR FACES, THEY ADD MORE PHYSICAL DESCRIPTIONS; WHEN THE SCENE CHANGES, THE ENVIRONMENT IS WRITTEN IN MORE DETAIL; WHEN THE SOUND IS UNSTABLE, THE REFERENCE AUDIO IS UPLOADED. BUT THE LONGER THE PROJECT, THE LONGER YOU FIND, THE LONGER THE HINTS..

ONE OF THE FIRST THINGS THAT MANY PEOPLE DO WHEN THEY MAKE AN AI VIDEO IS TO OPEN UP A BIG MODEL, DESCRIBE THE IDEAS IN THEIR HEAD BRIEFLY, AND THEN LET AI DIRECTLY GENERATE A HINT. IT SEEMS VERY EFFICIENT, BUT WHEN IT REALLY COMES TO BEING CREATED, IT IS OFTEN DISCOVERED THAT PEOPLE ARE NOT DOING THE RIGHT THING, THAT THE SPECIAL EFFECTS ARE NOT IN THE RIGHT SHAPE, AND THAT THE CAMERA RHYTHM IS NOT THE SAME AS THEY THOUGHT, AND THAT THEY CAN ONLY CHANGE THEIR TIPS AND REPEAT THEIR CARDS. THE PROBLEM DOES NOT NECESSARILY ARISE FROM MODELLING CAPABILITIES, BUT FROM THE VERY BEGINNING WE GAVE AI TOO MUCH "SPECIFY SPACE". IF AI DOESN'T EVEN KNOW WHAT YOU WANT, THEN EVEN IF IT WRITES A FEW HUNDRED WORDS AND EVEN A THOUSAND WORDS

NOW THE AI VIDEO CAN ALREADY PRODUCE VERY CLEAR FACES, NATURAL SKIN TEXTURES AND RELATIVELY COMPLEX PERSONALITY MOVEMENTS, BUT MANY OF THE WORKS STILL HAVE A PROBLEM THAT IS HARD TO IGNORE: TAKING ONE FRAME ALONE, THE PERSON SEEMS TO HAVE NO OBVIOUS ERROR, BUT ONCE THE PICTURE IS REALLY MOVING, THERE IS A CLEAR “AI SENSE”. THIS PERCEPTION DOES NOT NECESSARILY STEM FROM INADEQUATE MODELLING CAPACITY, NOR DOES IT NECESSARILY BE RESOLVED BY ADDING WORDS SUCH AS “8K, SUPER-HIGH RESOLUTION, FILM QUALITY, DELICATE SKIN”. MANY TIMES, WHAT REALLY AFFECTS A PERSON'S SENSE OF TRUTH IS WHETHER THE PERSON'S IDENTITY IS STABLE IN DIFFERENT LENSES AND WHERE HE IS LIVING..

MANY VIDEOS HAVE DONE WELL, BUT THE RESULTS ARE OFTEN LOST TO THE SOUND. IMAGES CAN BE MODELLED AND OPTIMIZED BY CLIPPINGS, BUT THE SOUND IS OFTEN SIMPLY TREATED AS “A SINGLE MUSIC” OR “A LITTLE ENVIRONMENTAL SOUND ENDS”. THE PROBLEM IS THAT WHEN MOST PEOPLE MAKE VIDEOS, THE SOUND IS THE LAST STEP, NOT PART OF THE DESIGN. WE DON'T GO THROUGH THE MUSIC LIBRARY AND THE SOUND LIBRARY MANUALLY, AND WE DON'T RELY ON FEELING TO FIND A PROPER BACKGROUND MUSIC. WE'RE GOING TO HAVE A MORE STRUCTURED WAY OF GETTING AI TO UNDERSTAND THE CONTENT AND RHYTHM OF THE WHOLE VIDEO FIRST, AND ON THAT BASIS..

If the person's skin is more natural, the lens is better, and the movement is more stable, why does the official focus on long narratives, multi-mode references and video editors? And why is the gap between 2.5 and 2.0 less obvious in some simple dance and animation scenes? It's really worth comparing, not “which version looks up to date”, but what really happened to them in their actual creation. Can the characters stay on? Can the state of the first shot continue to the next? Is the relationship between complex actions clearer? Can reference materials be used more accurately? Generating..

IN MAKING THE AI VIDEO, MANY PEOPLE WILL FOCUS MAINLY ON THE CONSISTENCY OF THE PERSON, THE QUALITY OF THE SCENE, THE FLOW OF ACTION AND THE PHYSICAL DETAILS. OF COURSE, THESE ARE IMPORTANT, BUT WHAT REALLY DETERMINES WHETHER THE AUDIENCE WILL CONTINUE TO LOOK AT THEM IS OFTEN THE FIRST FEW SECONDS. A LOT OF VIDEOS LOOK AT EACH FRAME ON THEIR OWN, BUT IT'S STILL FLAT. THE PROBLEM IS USUALLY NOT THE QUALITY OF THE PICTURE, BUT THE ABSENCE OF A MEMORY POINT AT THE BEGINNING. THE PERSON STANDS IN THE CENTER OF THE PICTURE, THE CAMERA MOVES SLOWLY, THE PERSON LOOKS UP TO THE CAMERA AND THE TITLE GOES DOWN. THERE WAS NO OBVIOUS ERROR IN THIS OPENING, BUT THE AUDIENCE WAS ABLE TO JUDGE AT FIRST SIGHT WHAT WOULD HAPPEN. THERE'S NOTHING LEFT TO CONFIRM..

THE HINT DESCRIBES WHAT IT WANTS AND THE CONDITION MAP DEFINES HOW THE PICTURE MUST BE GENERATED. LEARN DEPTH MAPS, GRAMMATICAL CHARTS AND DYNAMIC CONTOUR CHARTS, AND A SINGLE PICTURE CAN ALSO EXTEND SPACE, PERSONAL DETAILS AND PRODUCT ANIMATION. MANY PEOPLE USE AI TO GENERATE PICTURES OR VIDEOS, STILL FOCUSING THEIR MAIN EFFORTS ON HINTS: CONSTANTLY SUPPLEMENTING SCENERY, LIGHT, LENS, MATERIAL AND ACTION, HOPING THAT THE MODEL WILL FULLY UNDERSTAND ITS INTENTIONS. HOWEVER, WHEN THE TASK INVOLVES RECALCULATION, CHANGE OF LOCATION, MAINTENANCE OF A PERSON ' S STRUCTURE OR TRANSFORMATION OF A PRODUCT, TEXT ALONE IS OFTEN INSUFFICIENT. WORDS ARE GOOD AT DESCRIBING CONTENT, BUT THEY'RE NOT VERY GOOD AT SPECIFYING WHERE EACH REGION SHOULD BE..

FOR THE FIRST TIME, MANY PEOPLE USE AI TO MAKE SHORT FILMS, AND THEY FOCUS ON "IMAGES CAN BE SO FANTASTIC" THAT THEY FLOAT, EXPLODE, DEFORM, AND ADD A LOT OF PARTICLES, LIGHT AND FAST-TURNING. SUCH IMAGES MAY BE POPULAR, BUT THERE IS OFTEN A PROBLEM — THE AUDIENCE REMEMBERS THE EFFECTS, NOT THE PRODUCTS. THE REAL EFFECTIVE BRAND IDEA IS NOT TO PUT A STORY ON THE PRODUCT THAT HAS NOTHING TO DO WITH IT, BUT TO FIND A STORY FROM THE ORIGINAL FUNCTION OF THE PRODUCT. LIGHTING EQUIPMENT HAS CHANGED TO DARK, COFFEE MACHINES HAVE TURNED RAW MATERIALS INTO DRINKS, CLEANING TOOLS HAVE RESTORED CHAOS AND SOUND HAS CREATED RHYTHMS IN QUIET SPACES. JUST TAKE THIS CHANGE..

The AI video doesn't look real enough, it doesn't have to be a hint or a model that doesn't match the camera mission. The same frame, the same hint, is inserted into different models, and the results may be completely different. Some models keep people's skin, product material and image colours well, but they are not suitable for large-scale action; others are better at running, dancing, complex mirrors and multiple reference material, but the quality of the details generated is not necessarily what the project needs most. Choose Violin O3 or Seedance 2.0, depending on three aspects: how much action the lens takes, whether it relies on multi-modular reference..

When using reference maps to generate new images, the most common problem is not “models don't understand”, but “models look too much”. All you wanted to do was to use the position and position of the person in the reference map, but the model brought the color, the costume, the material, the light and even the wind. The more words are written, the more the requirements may compete for control. The key to solving such problems is to decompose reference material into different responsibilities: role maps for identity and clothing, style maps for colour and quality, depth maps for proximity, contours and space relations, and movements. Upon video generation, deep video can continue to move people..

IT'S HARD TO PRODUCE A FULL-SIZE-FITS-ALL VIDEO, AND OFTEN IT'S HARD NOT TO CREATE A GOOD-LOOKING SHOT, BUT TO MAKE A DOZEN OF THEM LOOK THE SAME. THE FIRST PART IS WARM AND THE SECOND ONE IS GREEN; THE SAME PERSON IS BLIND AND DARK; THE INTERIOR IMAGE IS SOFT AND CUTS OUTSIDE, AND SUDDENLY BECOMES SHARP. EACH PART ALONE IS NOT SO BAD, BUT IT'S ALL THE SAME. THE SOLUTION TO SUCH PROBLEMS CANNOT BE A LAST FILTER. A MORE CONSERVATIVE APPROACH IS TO CREATE COLOUR AND PHOTOGRAPHIC BENCHMARKS BEFORE GENERATION, TO REPEAT FIXED INFORMATION IN GENERATION, AND TO DO A UNIFORM CORRECTION AFTER GENERATION. THIS PAPER BREAKS THIS APPROACH INTO A DIRECT..

A LOT OF PEOPLE MAKE SHORT FILMS WITH AI, AND THEY GET USED TO STARTING WITH A NICE CHARACTER MAP, AN ATMOSPHERE, OR A "DO ME A REVERSE STORY." THIS LOOKS QUICK, BUT IT TENDS TO GET MORE AND MORE MESSY: PEOPLE LOOK LIKE DIFFERENT PEOPLE, THERE'S NO CAUSE OR EFFECT BETWEEN THE SCENES, AND THE CAMERA LOOKS GOOD, BUT DOESN'T KNOW WHY. THE PROBLEM IS NOT THAT THE PICTURE IS NOT SUFFICIENTLY PRECISE, BUT RATHER THAT IT DOES NOT FINISH THE STORY FIRST. THE SHORT SCRIPT IS NOT A BEAUTIFUL STORY SHEET, NOR IS IT A PILE OF CHARACTERS, SCENES AND REVERSES. IT'S A CONTINUOUS CHAIN OF CREATION: FIRST, AN IDEA, THEN A TARGET AND A CONFLICT FROM THE POINT; THEN, YES..

THE SAME IS TRUE OF THE AI VIDEO MODEL, WHERE PEOPLE GET SHORT CLIPS THAT ARE STABLE AND CONSISTENT, AND PEOPLE GET SHORT CLIPS THAT ARE FINE BUT IRRELEVANT. TOO OFTEN THE PROBLEM IS NOT JUST THE ABILITY OF THE MODEL, NOR THE LACK OF ADJECTIVES, BUT THE HINTS DO NOT TELL THE MODEL WHO IT IS, WHERE SHE IS, WHERE SHE STARTS, WHERE THE CAMERA STARTS, WHAT HAPPENS NEXT, AND WHERE THE MIRROR STOPS. A CONTROLLABLE AI VIDEO ALERT CAN BE DIVIDED INTO TWO LAYERS. THE FIRST LEVEL IS A GLOBAL SET-UP, WHICH IS RESPONSIBLE FOR MAINTAINING THE IDENTITY AND QUALITY OF THE ENTIRE WORK; AND THE SECOND LEVEL IS A SINGLE-PHOTO COMMAND, WHICH IS RESPONSIBLE FOR MAKING A VISIBLE MOVE IN THE CURRENT SECONDS. PLUS..

WHY ARE THE WORDS WRITTEN IN DETAIL, BUT THE RESULTS ARE UNSTABLE? THE PERSON IN THE LAST SECOND WAS IN THE OFFICE AREA, AND THE NEXT SECOND WAS LIKE A CHANGE OF SCENE; THE MESSAGE WAS “SCENES FOLLOW”, BUT THE PICTURE WAS TURNED INTO A FEW INCOHERENT IMAGES; THE RESULT WAS THAT THE CHARACTER HAD TO MOVE FROM THE DOOR TO THE TABLE, AND THE PERSON WAS NOT IN THE RIGHT POSITION, DIRECTION OR MOVEMENT. THE REASON IS NOT NECESSARILY THAT THE MODEL IS NOT GOOD, BUT THAT THE INPUT INFORMATION IS NOT “ENFORCEABLE”. AI UNDERSTANDS THE LANGUAGE, BUT WHEN THE TASK INVOLVES COMPLEX SPACE, MULTI-PERSON HIERARCHIES, CHASES, FIGHTS AND THE END OF A SINGLE LENS, IT CAN EASILY BE AMBIGUOUS. YOU SAID, "THE HOSTESS MOVES RIGHT THROUGH THE TABLE." AI NEEDS TO JUDGE..

PEOPLE CHANGE THEIR FACES, AND THEY THINK THEY'RE NOT VERY PRECISE. WHEN THE SCENE WAS IN DISARRAY, THE WORD “CONSISTENT” WAS ADDED TO THE INTRODUCTION. IF THE FIGHT ISN'T GOOD, THEN THE WORD “HEAT” “FILM SENSE” “SHOCK” BECOMES MORE FULL. THE COLOURS ARE NOT UNIFORM, AND AFTER THAT, ADD THE WORD "ADVANCED COLOURS". BUT THOSE WHO HAVE ACTUALLY DONE THE WHOLE PROJECT KNOW THAT THE HINT CAN ONLY SOLVE PART OF THE PROBLEM. IF THERE WERE NO ASSETS IN THE PRECEDING PERIOD, EACH IMAGE AND EVERY SHOT OF AI SEEMED TO BE RECHARGING. HE DOESN'T KNOW WHAT THIS CHARACTER LOOKS LIKE. HE DOESN'T KNOW IF THE SIDE IS THE SAME

WHEN A LOT OF PEOPLE DO THE A.I., THE FIRST REACTION IS TO GO WITH WORDS. IT'S LIKE A MOVIE, A SENSE OF EXCELLENCE, A SENSE OF TRUTH, A FEELING OF ATMOSPHERE, A FEELING OF GREATNESS. BUT YOU'LL FIND THAT THOSE WORDS ARE ALSO WRITTEN, AND SOME OF THEM ARE CREATED LIKE A MOVIE SCREENSHOT, SOME OF THEM ARE CREATED AS AN AL. IT DOESN'T HAVE TO BE A MODEL, IT DOESN'T HAVE TO BE A DRAWING, BUT YOU DIDN'T TELL AI WHERE IT SHOULD BE. THE LENS ANGLE DETERMINES THE RELATIONSHIP BETWEEN THE AUDIENCE AND THE PERSON. IT'LL PUT THE AUDIENCE AND THE PERSON AT THE SAME HEIGHT, IT'LL BE MORE REAL, IT'LL BE CLOSER; IT'LL MAKE THE PERSON LOOK AS IF HE OR SHE IS BEING OBSERVED, CRUSHED, FIT TO EXPRESS AND..

WHEN A LOT OF PEOPLE ARE DOING THE ROLE OF AI, THE FIRST REACTION IS TO FIND A GOOD REFERENCE, AND THEN MAKE IT HAPPEN. THIS APPROACH APPEARS TO BE FAST, BUT THE PROBLEM IS ALSO CLEAR: THE ROLE IS SO EASY TO BE LIKE OTHERS, WITHOUT ITS OWN VISIBILITY, AND IT IS DIFFICULT TO STABILIZE WHEN IT COMES TO CONSISTENCY. MORE IMPORTANTLY, THE RISK OF COPYRIGHT AND PORTRAITS MAY ALSO ARISE IF REFERENCE MAPS ARE DRAWN FROM REAL PEOPLE, VIDEO CHARACTERS OR WEB-RED PHOTOGRAPHS. THE REALLY APPROPRIATE METHOD FOR LONG-TERM USE IS NOT "REPLICATE A FACE" BUT "DESIGN A FACE". THE CORE OF AI'S FACE IS NOT TO MAKE THE FIVE OFFICIALS MORE BEAUTIFUL, BUT TO MAKE THE CHARACTER'S FACE STRUCTURAL, STENCH, MEMORY, AND..

A LOT OF PEOPLE DO AI VIDEOS WITH THEIR FULL FOCUS ON CHARACTERS, COSTUMES, SCENES AND STYLES. IT'S A LONG TEXT AND A FINE PART, BUT IT'S STILL LIKE "GENERATING A PICTURE" WHEN IT COMES OUT, AND IT'S LACKING THE QUALITY OF THE REAL IMAGE. OFTEN THE PROBLEM IS NOT THAT THE PERSON IS NOT GOOD ENOUGH OR COMPLEX, BUT THAT THE LIGHT IS NOT DESIGNED. IN THE REAL PICTURE, THE LIGHT IS NOT JUST LIGHTING PEOPLE, IT DETERMINES THE RELATIONSHIP BETWEEN THE SUBJECT AND THE BACKGROUND, WHETHER THE PERSON IS EMOTIONAL, WHETHER THE PICTURE IS HIERARCHICAL OR WHETHER THE AUDIENCE BELIEVES IT EXISTS. SO THIS SECTION, WE DON'T SPEAK COMPLICATED STYLES, NOR..

THE HARDEST PART IS TO PRODUCE AN AI FIGHT VIDEO, NOT TO COPY IT INTO A VIDEO TOOL, BUT TO GET A REALLY USEFUL ONE. IF ONLY AI SAID “WRITING AN IRRITATING PUNCHWORD”, IT DID NOT KNOW WHAT THE TWO ROLES LOOKED LIKE, WHAT WEAPONS WERE USED, WHERE TO FIGHT, OR WHAT ACTION, EFFECTS AND LENS SHOULD REFER TO. THE FINAL CONTENT MAY BE POPULAR BUT CANNOT MATCH THE ROLES AND MATERIALS IN HAND. A MORE STABLE APPROACH WOULD BE TO GIVE ALL CHARACTERS, SCENES, ACTION MOVES, MIRRORS AND SPECIAL EFFECTS REFERENCES TO AI IN SUPPORT OF THE MAP, AND TO USE A CLEARLY NUMBERED DIRECTIVE TO MAKE IT..