Facial features are the most afraid of fake, rigid and face change. This oneAI TutorialsFrom a human face map, it's clear how to predetermine it and then make it happen。
Then, in my thinking and in my steps, you can do something like this
seedance2.5/seedread 5.0pro
IF YOU DON'T KNOW HOW TO USE THE MODEL, YOU CAN REFER TO THE AI ACADEMYhttps://www.1ai.net/56824.htmlUse it with the curriculum
First, get the initial picture ready
When making facial feature videos, the initial picture basically determines the quality of the final picture。
If the skin, eyes, hair and makeup in the picture are not real enough, it is difficult to fully remedy the subsequent generation of videos. Especially when the face is almost full
Faces, five officials and skin problems are magnified。
The first step, therefore, is not to rush into making a video, but to get a real, delicate, natural image of the person's face。
This map better meet several conditions:
The face is almost full;
The head is naturally cut off by the image;
Only a small number of necks appear at the bottom;
The face is bright, but not exposed;
Hair, eye, lipstick and overall makeup are clear;
(a) Clean and defunct, but without light;
Relax, don't stare, don't smile。
ii. First look for facials and then extract the five profiles
DIRECTLY ALLOWING AI TO CREATE A "PRETTY PERSON" IN A VACUUM, THE RESULT IS OFTEN RANDOM AND THE PERSON IS PRONE TO LACK OF IDENTIFICATION. MY APPROACH IS TO START BY LOOKING FOR A PROFILE OF MY OWN. BUT THIS FIGURE WILL NOT BE USED DIRECTLY IN THE DRAWINGS, BUT ONLY TO HELP AI OBSERVE AND EXTRACT THE FIVE OFFICIAL FEATURES。
this is where you can look for pictures from a little red book/pinterest platform
THE PURPOSE OF THIS IS TO TRANSLATE THE FACE, EYEBROW, NOSE, LIPS, HAIR AND MAKEUP IN THE REFERENCE DIAGRAM INTO WORDS AND TO CREATE A NEW AI FIGURE THROUGH A PURE GRAPHIC。
YOU CAN PASS THE REFERENCE TO AI AND USE THE FOLLOWING:
I HAVE A PROFILE OF THE PERSON, AND I'M GOING TO MAKE A NEW ONE. PLEASE ANALYSE AND EXTRACT THE FACE, EYEBROW, EYEBROW, NOSE, LIPS, COLOUR, HAIR AND MAKEUP FEATURES VISIBLE IN THE PICTURE, AND MAKE UP A CHINESE TEXT MESSAGE THAT CAN BE USED DIRECTLY IN A PURELY LITERARY DRAWING. THIS REFERENCE CHART WILL NOT BE USED FOR THE OFFICIAL PRODUCTION OF THE PICTURE, SO THAT THE FIVE OFFICIAL FEATURES WILL NEED TO BE DESCRIBED IN A SPECIFIC, ACCURATE MANNER AND IN A MANNER CONSISTENT WITH MAKEUP AND HAIR. THE DRAWINGS ARE WRITTEN ON THE FACE OF THE ULTRA-NEAR, WITH ALMOST FULL FACE, WITH THE HEAD CLEARLY PAINTED, THE TOP HAIR CIRCUMCISED, AND ONLY A SMALL NUMBER OF NECKS ON THE BOTTOM, WITHOUT THE SHOULDER. THE PERSON'S FACE IS BRIGHT, NATURAL, DELICATE, BUT EXPOSED; HIS BACKGROUND IS CLEAN AND HE HAS A SLIGHT SHALLOW LANDSCAPE, BUT HE DOES NOT APPEAR IN THE HALO, LIGHT SPOTS, CONTOURS OR LIGHT BACKGROUND. THE PERSON'S FACE IS LAX. FINALLY, COMPLETE AND REVERSE REFERENCES ARE PROVIDED。
THE MOST IMPORTANT THING HERE IS TO TELL AI THAT WHAT WE NEED IS "TEXTUAL FEATURES THAT CAN BE USED FOR PURE GRAPHICS" RATHER THAN MAKING IT A DIRECT REFERENCE。
III. PRODUCTION OF INSTITUTIONAL TEXTS
When you get a organized facial hint, you can go to the texture phase。
I'm using this dream Seedream 5.0 Pro. In this comparison of the same hint, the person it produces is more in line with my expectations in terms of five officials, skin and overall precision。
I've tried Image 2.0 and NanoBanana Pro, but in this case, the real sense of the results and the precision of the person did not work as well. This is a comparison between the current narrative and the type of person, and does not mean that all missions will have the same conclusion。
iv. Don't prepare too many emoticons
At the beginning, I tried to prepare a reference map of different expressions, and I wanted the model to do it in sequence。
As a result of the actual tests, there are two problems that can be caused by too many emoticons:
First, people can change their faces too quickly, like they're changing pictures on a case-by-case basis; second, their eyebrows, eyes and mouths can easily change at the same time and appear more mechanical。
FINALLY, I REDUCED THE EXPRESSION REFERENCE AND WENT ON TO USE ONLY ONE INITIAL MAP, SO THAT A WOULD BE FREE TO PRODUCE INTERMEDIATE EXPRESSION BASED ON MOVEMENT AND EMOTIONAL LOGIC。
The reference diagrams are for targeting people and the hints for arranging performances. Such changes in sight, blinking, turning and expression are usually more consistent。
V. LET AI EXTRAPOLATE VIDEO ALERTS
Here's the generic hint I'm using
ONCE YOU GET A SATISFACTORY INITIAL MAP, ALLOW AI TO WRITE VIDEO ALERTS BASED ON PICTURES, TIME AND KEY ACTIONS。
You're a good manMicro-expressionTHE DIRECTOR OF THE SHOW MOVEMENT AND THE PRODUCTION OF VIDEO BY AI. PLEASE EXTRAPOLATE A NATURAL, CONSISTENT, DIRECT-USE CHINESE TEXT MESSAGE FOR THE DRAWING VIDEO MODEL, BASED ON THE PERSONAL REFERENCE MAPS, VIDEO TIME AND KEY EVENTS I HAVE PROVIDED。
My message:
– Person reference diagram: [upload pictures or file path]
– Video duration: [insert seconds]
– Critical event:
– Event time: [insert time point]
– Other requirements: [Completion of design, light, background, lens or emotional requirements]
Please design this as follows:
1. The initial expression of the person in the reference map, the direction of the sight, the head angle, the structure, the hair, the makeup, the light and the background are first identified. The reference chart is a frame of reference for the physical appearance of the person and does not redesign the person。
2. Designing the entire video as a continuous performance with a causal link, e.g., an initial state of detection of a change in attention transfer event, an instinctive response to the recovery of final emotions. Do not add to the image for no reason of surprise, chirping, angry or smiling。
3. The reference expression is used only as an emotional anchor and does not require a definitive reset. Natural meso expression, change of sight, blinking, breathing, slight turning of head, bowing of head, lifting of the chin and physical stress reaction outside the reference map are allowed。
4. Emoticons cannot change mechanically. Eyes, eyelids, eyebrows, heads, cheeks and mouthbrows shall be activated in sequence. For example, the eye detects, the head is slow to follow; the eyebrow reacts first and the mouth is changed later。
5. Each emotional change shall include: preparation for action, formation of expression, short symphony and connection to the next state. The speed of movement has been slow and slow, and the real response has been delayed and the slight inequity has been observed。
6. If hand contact, touch, squeeze or pull is present, describe in detail the direction of the hand ' s entry, the track of movement, the time of contact, the position of the finger, skin stress, the deformation of the cheek, the head, the echoes after release and the subsequent reaction of the person. The actions must be consistent with the true physical order。
7. Targeted identity: five officials, face, age, colour, hair, eyes, lipstick and makeup remain unchanged. Lock the lens, the image, the exposure, the light and the background, unless I specifically ask for change。
8. Rational rhythm based on the total video time. It is not appropriate to have too many expressions, and each major emotion must have sufficient time to form and stay. Do not even divide multiple expressions into mechanical time nodes。
Please output as follows:
The design basis summarizes the initial state, the emotional trigger point, the reaction chain and the final state in short language, without lengthy analysis。
ii. The full video message contains:
Images of people, images, lenses, drawings, light and background。
[Current action time line] describes sight, head, eyebrow, mouth angle, breath, interactive action and emotional development by specific seconds。
The emphasis is on priority, late response, minor asymmetries, transitional expression, stress feedback and emotional symmetry。
[Inverted hint] Excludes face change, five drifts, mechanical face change, all five change at the same time, no reason-to-face, mouth speech, facial fluidization, finger malformation, lens disorder, exposure flash and background change。
The final reminder must be specific, continuous and have an enforceable time line, but not be written in a frame-by-frame description; it focuses on a sequence of dynamic processes with causality。
There are also key video text messages at the top, which you can use with minor modifications:
Using the reference map as the reference figure for the only person, only the expression is changed and the face, five officials, colour, hair, eyeline and rose-paste is maintained. It is written on the face of a close distance, with the person looking in the face and almost full of images; the head is clearly painted, the top hair is circumcised, the left and right parts of the head are painted, and only a small number of necks are shown on the bottom and no shoulder appears. Cameras are paralleled with eyes, 100 mm human lenses, fixed position, fixed focal length, focus locked eyes; no pull, no shake, no camera shake. Large areas of soft light on the front side are evenly luminous, bright but unexposed, and the light and exposure remain stable. Keep the background clean, low saturated and light pink, without light, without radium, light, dispersion, reverse and contours. The 4K super-synthetic body of the person, with a clear natural skin, eyelashes, lips and hair, and a stable continuous frame, with no scintillation, after-effects, liquidation or five official drifts。
The video message reads as follows:
0-1.2 seconds: The person does not look at the lens, the head is slightly tilted to one side, the view naturally falls below the side of the lens and the facial muscles are completely relaxed. Keep a slight breath and the eyeball produces a small natural movement。
1.2-1.5 seconds: The person is perceptive of the lens, with only a quick eye glimpse of the lens and a brief contact with the lens; the head is not moving for a while, the sight is then moved naturally and the face is not visible。
1.5-2.5 seconds: The person blinks lightly, and the view stays by the lens for a moment. The lips remain relaxed, the head floats millimetres with the breath, and the spirit is quiet and natural。
2.5-3.8 seconds: The sight is slowly re-moved and the head is slow and light. The eyes go through the eye, the head follows, and eventually the lens is seen naturally。
3.8-5.8 seconds: The person is silent and eye-to-eye, does not deliberately smile and does not make a specific expression. A natural blinking occurs during the period, with a slight focus change in the eye; the lips are sometimes tightened and relaxed, and the chin produces small motion with the breath。
5.8-6.8 seconds: The person's residual light detects something on the right side of the picture, the pupils move slightly to the right, and the eyebrows on the side are raised only in a very small way, showing natural care and doubt, and do not suddenly surprise or stare。
In the 7th second, a real, natural woman's hand reaches right in. Fingers are fine, five fingers are structurally correct, short and clean natural fingernails are not worn。
7.3 seconds, the thumb and index finger gently squeeze a small piece of soft meat on the right cheek of the person. The cheeks are naturally dented with pressure, the soft flesh around them is slightly drummed, and the mouth angles produce real asymmetrical tractions。
The person was held up briefly, with the soft and soft intuitive lifting of the eye on one side and the small opening of the other eye; the eyebrow produced a small reaction, the lips were softly held, and the head moved a few millimetres towards the other. Reactions are naturally restrained, as if they were suddenly squeezed by a familiar person, rather than exaggerating。
8.1 seconds before the finger was released, the cheek softly rebuffed and the hand pulled out of the picture 8.5 seconds ago. The person blinked once and showed a short period of impotence, followed by one side of the mouth and one of the other side slowly followed。
8.5-10 seconds: The person gradually develops a soft, loose smile and the eye returns to the lens. Finally, it breathes gently and blinks naturally, and the smiles are mainly in the eyes and mouth。
In the first seven seconds, there were no obvious surprises, lulls, angers or frustrations, but only real vision changes, breaths, blinks and subtle deities. The person does not speak, has no mouth, does not mechanically mutilate the face, liquidizes the face, deformed the finger, moves the lens or blinks the light。
VI. Transfer of pictures and tips to video models
Finally, the initial picture and the resulting video alert are uploaded to Feedance 2.5, setting the required video time and starting to be generated. (Recommends Mr. 480p to adjust and eventually produce 1080p, which is completely different.)
