The one from the previous timeAI skit tutorial,INTRODUCTION TO THE AI SHORT PLAY, FROM 0 TO 1
A number of small partners say it's too dry, and we continue today to share the rest。
The main content is video generation, spectroscopy optimization and sound design, because editing still requires more systematic learning, multi-basketing, and we've got this, at least one that looks goodAI skits.
catalogs
1.Video Generation
- Introduction to Video Generation Tool
- Video Generation Mode
- Full Reference 2.0 Usage Techniques
2.Mirror Optimization
- Meaning of the specs
- Recommendations for the outchart adaptation tool
- Speculation technique
- Optimizing the lens and image remediation
3.Sound Design
- music
- soundscape
- Phrase (not necessary)
V. Video generation
Due to the number of tools, it is not possible to describe them in detail. In response to the current AI Human Shortplay, we focus on the Seedance 2.0 model from the perspective of a combination of effects and the use of thresholds。
i) Introduction to Video Generation Tool
Syndication Platform - LibTV
Double-click the canvas to create video nodes. In the model selection column here, you can see many video models. If there's a lot of models, and you're confused, then focus on the Feedance series and the Kling series。
Core advantages: In addition to the blood-filled version of Feedance 2.0, many other video models can be used. There are no limits to real people。
Network of Video Model Platforms
Seedance 2.0: A very high-ceiling "one-clicked" device
Core strengths: Official platform for Feedance 2.0。
LIMITATIONS AND PAINS: IF THE VIP MODEL IS NOT USED, THE MATH IS VERY STRESSFUL AND THE PEAK PERIOD IS EXTREMELY LONG, BUT THE VIP MODEL IS MORE EXPENSIVE. THE SECURITY AGREEMENT AT THE BOTTOM OF THE PLATFORM IMPOSES EXTREMELY STRICT RESTRICTIONS ON THE PRODUCTION OF REALITY-WRITING FILMS, WHICH CAN EASILY TRIGGER A PORTRAIT OR ACTION AUDIT INTERCEPTION。
Spirit: Hexagonal warriors, with the same quality and stability
Core advantages: Very solid in the combined effect of raw video. Before the launch of Seedance 2.0 in the dream, it was the absolute priority of the closed source model. It is extremely clear and stable and supports raw 4k resolution video。
Limitations and pains: It's more of a general performance for extremely difficult large dynamic movements or extreme mirrors. It's not like Seedance, it's brainless, it's a great test of the author's own audio-visual linguism and the precision of the hint. The accuracy of the characters ' lines is good and bad, with a higher rate of long-screech materials over 10 seconds. The latest videos 3.0 and 3.0 Omni models are more expensive and costly to try。
(ii) Video generation mode
AI VideoThere are three main modes of use: Vincent's video, Tusheng's video and reference-based video (full reference/photogram reference)。
Given that the Full Reference Model of Seedance 2.0 has demonstrated a great deal of effectiveness in the creation of the AIMS short play, we would like to highlight the use of this model to produce an AIMS short play。
(iii) Feedance 2.0 Full Reference Model Usage Techniques
There are many ways to write a hint, and there are no so-called standard hints, and the following is recommended if a short story is to be completed。
Complex dispatch segment (recommended): asset + spectrophs
IF YOU ARE THE DIRECTOR YOURSELF, OR HAVE A CERTAIN AUDIO-VISUAL LANGUAGE BASE, THERE ARE VERY FINELY DETAILED REQUIREMENTS FOR THE INTERNAL SCHEDULING AND SWAGGERING OF THESE 10-SECOND SEGMENTS (E.G., YOU HAVE TO CLOSE YOUR FINGER, THEN YOU HAVE TO CUT YOUR FACE, AND THEN YOU HAVE TO CUT TO A TWO-PERSON PANORAMA, ETC.), IT IS NOT ENOUGH TO FEED THE SCRIPT FOR AI TO BE FREE。
IN THE CURRENT FAST-PACED AI SHORT-DRAMATIC WORKFLOW ENVIRONMENT, IT IS UNLIKELY THAT THE WHOLE PROCESS WILL BE ABLE TO PRODUCE A FIRST-OUT SPECTROSCOPY, FOLLOWED BY A VIDEO, BUT RATHER TO LEARN TO CONVERT THE SCRIPT INTO A SPECTROSCOPY HINT THAT WILL GUIDE THE MODEL TO PRODUCE MORE ACCURATE VIDEOS, WITH A FEW SECONDS TO A FEW SECONDS。
Phrasing formula:
Handheld (fixed roles, scenes, sound and negative hints, spatial position relationship description) + Spectrum tip
The uploading of characters, scenes, audio, etc., can be carried out for a lot of in-depth learning, for a temporary reason, and you can focus on the courses of our end-of-text microshop。
Note: Space relations need not be described if role position maps are prepared in advance and provided as a reference for Feedance 2.0。
WE DO NOT NEED TO UPLOAD THE SOUND AS A REFERENCE FOR THE FIRST PLAY FOR EACH ROLE. THE VOICE OF THE ROLE WILL BE EXTRACTED AFTER THE VIDEO IS GENERATED BY AI. AS A RESULT, THE NEXT VIDEO CAN BE UPLOADED AND USED, FOLLOWED BY A DETAILED PRESENTATION。
Spectrum phrase: The first version of the Spectrum phrase can be written using the following phrase template。
YOU'RE A SENIOR FILM DIRECTOR AND PROFESSIONAL AI VIDEO-GENERATED HINT ENGINEER. YOUR TASK IS TO TRANSLATE THE USER'S "SCRIPT TEXT" OR "SCRIPT SCRIPTS" INTO STRUCTURED VIDEO ALERTS THAT CAN BE USED DIRECTLY IN THE AI VIDEO GENERATION TOOL. YOU NEED TO HAVE A VERY STRONG AUDIO-VISUAL RHYTHM AND EDITING MIND。
** [core dismantling rules]** (Please strictly comply with all of the following):
Paragraph length and cutting:
THE TOTAL LENGTH OF THE HINT FOR EACH INPUT IS APPROXIMATELY 13-15 SECONDS (BASED ON 15 SECONDS AND FINE-TUNED TO THE RHYTHM OF THE ACTION OR LINE). DIFFERENT SCENES OR EMOTIONAL TURNING POINTS HAVE TO BE CUT INTO DIFFERENT MATERIAL PARAGRAPHS, EACH OF WHICH IS SEPARATED BY “-” AND “[PART X]”。
The lines are absolutely faithful and formatted:
The line words cannot be altered and must be retained without any change. Do not use a hard "word:" label to take place. Please include the lines directly in the image description in the form of a particular phrase, “the content of the lines”。
Sound Properties Distinction:
Strictly separates and labels the origin properties of the sound. The words in person are marked as: “...”. The words in the heart are marked as: "...." A person who is not in the picture but whose voice appears (or cuts into the other person's reaction lens) is marked as: the voice of a particular painting: “...”。
Short play fast-paced clips (reaction lens mechanism):
If the character's lines are longer, it is absolutely impossible to stare at people with a long shot. It is necessary to enter the reactions of the present audience (such as shock, greed, fear, anger) in the middle of the tempo. When cutting into the reaction lens, the original line cannot be broken and the voice of the speaker should be marked。
Dialogue vision and space logic:
The dialogue scene must be based on a standard back-to-back set. Before a violent encounter, or when the scene is just switched, the space lens (e.g. architectural appearance, visual feature) can be usefully used to guide the sound of the line and to promote the film。
Sound processing:
Global setting: Only sound, not music. Do not produce any subtitles. Only clear environmental/actional sound (e.g. footsteps, cracks, breathing) is written at the end of the reminder. If the lens is not clearly sounded, the sound field is simply omitted and is not written on purpose。
** [Standard Output Format Template]**
[Reflection Paragraph 1]
To present a [time, environment, atmosphere] [core event] scene using [replacement to a user-designated visual style]。
Only sound, not music. Do not produce any subtitles。
0-X SECONDS, [SEGMENT], [CARGO/FIXED], [DETAILED DESCRIPTION OF THE CONTENT OF THE IMAGE, INCLUDING THE MOVEMENT OF THE PERSON, THE STATE AND THE STRUCTURE, ETC.]. [FACENAME] SAYS/INNER MONOLOGUES/PEATHINGS: “[ONE WORD]”. SOUND: [SPECIFIC SOUND DESCRIPTION, NONE WRITTEN DIRECTLY]。
X-X-SECONDS, [SEGMENT (MARKED INVERSE/REACTION LENS)], [CARGO/FIXED], [DETAILED DESCRIPTION OF THE CONTENT OF THE IMAGE]..
[Reflection paragraph 2]
To present a [time, environment, atmosphere] [core event] scene using [replacement to a user-designated visual style]。
Only sound, not music. Do not produce any subtitles。
0-X SECONDS, [SPEECH]..
...
** [Input area]**
Please address the following in accordance with the above-mentioned rules and formats:
VISUAL STYLE: (PLEASE FILL IN HERE THE STYLE YOU WANT, E. G. REAL-TIME VISUAL STYLE / 2D / 3D ANIMATION STYLE, ETC. IF YOU DON'T FILL IT IN, YOU'LL BE ABLE TO USE A REAL VISUAL STYLE
Script:
"Play your script here or a lens table"
Note: The generated video spectrophs are only the first edition and require manual modification based on the effects of running video generation。
Access sound assets
IMPORTS THE GENERATED VIDEO INTO THE CLIP TO FIND THE AREA WHERE THE CORRESPONDING CHARACTERS SPEAK. SHORTCUT I SETS THE STARTING POSITION, AND O SETS THE END POINT。
Select a clip from which the different players speak, with no more than five seconds suggested。
the audio is then exported as a sound asset for the role. the audio assets of the corresponding role can then be used repeatedly as audio references to generate a video with a copyance 2.0。
Material connection techniques
The short play is co-mingled by several pieces, and it is difficult to access it at a later stage if the material generated is photocopied and the positions vary considerably。
1) Frame relay
Uploads the last frame screenshot from the previous video as the frame reference for the next video. This ensures, to the greatest extent possible, a seamless connection between the space, light and character of the two lenses at the time of editing。
in the case of video material generated in libtv, the mouse is suspended on the video and there is a camera icon at the lower right corner that can cut off the head or tail frame。
EXTRACT THE FRAME OF THE A-SECTION MATERIAL
You can also download video material and import it into a cut or any other clip software to extract it:
AND WHEN YOU GO ON WITH THE NEXT TEXT, WRITE: @PAGEX AS THE FRAME OF REFERENCE
THE TAIL FRAME OF THE A SECTION MATERIAL IS THE FIRST FRAME OF THE B SECTION MATERIAL
2) Extension of Video
You can also upload a clip and then request the model to extend the video in the hint. In this way, the model can refer directly to the frame of the previous video, but also to the sound used in the previous video, so that there is no need to extract the sound in advance as a sound reference。
given the current limitations of the platform on the face of real people, this method is almost impossible to continue in the event of a dream or other platform. in libtv, however, it is still only necessary to follow the method described in the previous session, which will require an extension of video material to perform a follow-up check-up check-up check-up of the fooddance 2.0, which will be completed after review。
ALL YOU HAVE TO DO IS START WITH THE HINT: EXTEND THE VIDEO BY XX SECONDS。
However, it was measured that the rate of the cards was a little high, that the effect was not as steady as the frame of the last video, and that the reduction of the sound was more general。
In addition to the two methods described above, the interface between the materials also addresses several basic principles of lens assembly. These require us to accumulate in the field, and the next class will elaborate on this。
It's like a dream
if a small partner is created using libtv's feedance 2.0 model, this part of the text below can be omitted。
for small partners using the soedance 2.0 model of other platforms, at present, when uploading reference maps, the dream platform places restrictions on the photographs and faces of real people and intercepts extremely authentic photographs or star faces. if your script is to use a particular physical appearance, the probability of directly uploading the original image will be subject to failure or violation。
HOWEVER, THE PRODUCTION OF A HUMAN-IMAGE SHORT PLAY IS ABSOLUTELY UNWIELDY OF REAL-LIFE CHARACTER ASSETS. HERE ARE A FEW WAYS TO IMPROVE THE SUCCESS RATE。
a. local face rotation (possibly change in face profile)
Redrawing of the face of the picture using photo models such as dream/banana。
Phrasing: Only the skin area (face) will be transposed to a colored sketch effect, and everything else includes clothing, hair color, etc。
This would allow the face to be just below the system's threshold for determining “absolutely real photographs”, while preserving the human bone and core features. The method was almost 100 per cent successful, and only face redrawing was done, and clothing details remained at a high level of reduction. However, the disadvantage is that when video material is generated, the facial similarity is slightly reduced。
b. in-house wash maps of platform stations (required several times)
This is a very practical and clever process of over-hearing, and the core logic is to put the Platform ' s trust in the data produced by the home。
THE THREE VIEWS OF THE HIGH-PROFILE PERSON THAT YOU'RE PREPARED FOR, EITHER UPLOAD IT TO THE IMAGE-GENERATION MODULE OF THE DREAM FIRST FOR A MINOR CHANGE, OR DIRECTLY USE THE HINT TO GIVE THE DREAM A LINE TO THE CORNER OF THE PICTURE: “THIS PERSON IS CREATED BY AI”。
AFTER THE CHANGE IS COMPLETED, YOU DRAG THIS IMAGE INTO THE VIDEO GENERATION MODULE DIRECTLY FROM THE DREAM DESK. THIS MUST BE DONE INSIDE THE DREAM PLATFORM! THE PLATFORM SYSTEM HAS THE PROBABILITY OF IDENTIFYING IT AS A LEGITIMATE PRIMARY ASSET GENERATED BY ITS OWN AI. DON'T PLAY SMART IN LOCAL P.S. AND THEN UPLOAD IT. IT'S NOT WORKING。
The advantage of this approach is that the role face is better recovered, but there is still the possibility of failure. After producing four graphs at a time, drag them into the video generation module below as a reference map and try them one by one, which is usually available。
For example, when we took the first video, the scene of the male-female co-sitting was taken as a case in which, with the addition of text at the top right corner, only the third picture was found to bypass the review。
c. face plus grid (face grid requires multiple cards)
For the four-view character diagrams (typographical + triangulation), the main restriction is the left-face feature, either not four-view, only three-view, not one on the left, which is largely successful, but the resulting video can easily become a dream-face as soon as it reaches the immediate view of the person。
In order to address this problem, it is only necessary to close the face on the left and add a grid that will ensure both a basic face profile and some covert treatment, with a success rate of about 70 per cent. Nano Banana 2 would be better adapted (see the latter chapter using the method), i.e., the dreaming face changes, with the following hint:
d. facial close-ups
Essentially, if only three viewes (without face features) would hardly trigger the platform's face-check limits. The key to the problem is the face close-up in the four views. But if you don't close your face, only three views are left, and once the video is made, it is likely that it will become a “dream face”。
SO WE CAN DRAW A SMALL BOX TO COVER THE SIDE OF THE FACE IN THE PS OR OTHER GRAPHICS TOOLS, AS SHOWN BELOW, SO THAT WE CAN BASICALLY PASS THE AUDIT. WHEN THE VIDEO WAS RUNNING, AI WAS ABLE TO COMPLETE THE MINISTER'S FACE THROUGH THE OTHER HALF OF HIS BRAIN, BECAUSE IT COVERED ONLY A SMALL HALF OF THE AREA。
e. multiple role generation failure
For multi-player video material, don't come up and refer to all the characters like this, and then run the 15-second video。
It is important that one character try to use the two methods above, so that each role can run a four-second video, and that each role map can be reviewed before multiple roles can be combined to produce material。
Seedance 2.0
When short-screen dramas are produced using Feedance 2.0, large numbers of original subtitles are generated without control. These subtitles are often muffled and the format of each subtitle varies. Such subtitles must be removed, as they will certainly not be accepted by A, and will need to be removed at a later stage by adding their own subtitles to the editing software。
1) CLIPSHOT AI ELIMINATION
IT'S RECOMMENDED TO USE THE AI ELIMINATION FUNCTION IN THE CLIPS, SELECT THE SMART SELECTION, FIND THE LONGEST FRAME OF SUBTITLES IN THE MATERIAL, RUB THEM, AUTOMATICALLY RECOGNIZE SUBTITLES, AND THEN REMOVE THEM. THIS APPROACH REQUIRES THE CONSUMPTION OF THE CUT-OFF POINTS OF THE MEMBERS, BUT IT IS CONVENIENT TO OPERATE DIRECTLY IN THE CUT-OFF SOFTWARE, NOT FOR RETROFITTING, AND IS MORE EFFECTIVE。
2) Libtv Smart Subtitle
You can select both modes of smart erase or check box erase。
SMART ERASE MEANS THAT AI AUTOMATICALLY RECOGNIZES SUBTITLES AND REMOVES THEM WITH A SINGLE KEY, WHICH IS SLOWER。
IN THE BOX ERASER MODE, THE LINE WITH THE LONGEST SUBTITLES IS SELECTED AND THE AREA OF SUBTITLES IS SELECTED MANUALLY. SUBSEQUENTLY, AI WILL WIPE THE REGION AS A WHOLE. THIS WAY IS RELATIVELY FAST。
VI. Mirror optimization
(i) Meaning of the spectrograph
In the old era of hand-flowing, the spectroscopy had greater significance than anything, and its pros and cons directly determined the overall rhythm, image aesthetics and consistency of the video. The spectrographer needs to carefully plan the scenery, the image and the angle of the film, and to break it down into a very fine picture in a coherent and complex way。
But now, with the emergence of large original video models, such as Seedance 2.0, they can come in for more than 10 seconds, with complete multi-space-to-switch material, provided that they are given script tips and assets。
But that doesn't mean that it doesn't need any more. At present, the meaning of the spectroscopy has been changed from a flow of water to a precise surgical operation, which is designed to address three main pain points:
1. Reduction of video-draw rate: For extremely complex maximum lenses, the success rate can be greatly enhanced by a precise static spectroscopy, followed by a run-off video。
The end-of-the-pipe remediation of the residual lens: The long materials that move for a few seconds are partially damaged (e.g., the last few seconds when the face collapses) and optimized by "Scratches and recharacterization"。
3. Empty lens and emotional fixation: The solution to the big video model is hard to get out of the statistic close-up and emotional emptiness (e.g. raindrop close-ups, letters close-up, people tearing down)。
(ii) Recommendations for the charting tool
Friends who play AI paint know that the main force of our drawings in the last two years was Midjourney and Stable Diffusion。
Midjourney is a good aesthetic, but requires a scientific access, a high fee threshold, and the understanding of Chinese and Chinese Eastern elements is often “sculptive”; and Stable Diffusion WebUI and ComfyUI workstreams, although free of charge and with greater control over images, have a certain threshold for starters, the debugging process is cumbersome and can slow down in the current short-time, fast commercial production process。
In our short play visual production process, there are lots of graphic platforms. Considering the characteristics of the model itself, we recommend the following models:
Nano Banana Series (Nano Banana 2 and Nano Banana pro)
GPT Image 2
Seedream series (i.e. dream/peasette photo model)
Recommended platform: LibTV
A large number of photo and video modelling tools were accessed。
The Lib Nano series model is equivalent to the Nano Banana series model, which is equivalent to the GPT Image 2 model。
(iii) Spectrograph technique
Standard Graph Phrasing Formula
IN THE PRODUCTION OF A REALITY PLAY, THE BASIC HINTS OF THE AI SCRIPT CAN BE FOUND IN THE FOLLOWING FORMULA:
[articular style/media] + [spectrum and perspective] + [subject description] + [environment scene] + [photo and tone] + [drawing and quality]
OF COURSE, IT IS NOT THE ONLY ONE THAT IS WRITTEN AND FORMATTED. IN THE COURSE OF LEARNING TO PAINT AI, YOU CAN EXPLORE A VARIETY OF HINTS. BUT WHATEVER THE FORM CHANGES, THE ESSENCE REMAINS THE SAME. THE KEY TO GENERATING PICTURES THAT FIT THEIR NEEDS IS TO CAPTURE THE CORE POINTS. FRANKLY, IN THE TEXTUAL FORMULA, ONLY THE SUBJECT DESCRIPTION AND THE ENVIRONMENT SCENE ARE THE MOST CRITICAL, WHILE THE REMAINING ELEMENTS ARE MERELY THE ADDED VALUE OF SUPPORTING MODELS FOR BETTER COMPLIANCE WITH INSTRUCTIONS。
For example:
- Wang Jia Wei's film sense, reality film. Nearview, 45 degrees. A 30-year-old man with a hard face, dressed in a black custom suit, with a little head down and a silver lighter in his right hand. The background is the luxurious Office of the Managing Director, with a vague urban night view outside the window. Rembrandt, the cold color is the main, and the lighter flames provide warm and colourful light. Real skin pore texture, shallow view, film pellets。
Multi-Asset Combination Biograph (reference + hint)
After the role and scene assets have been established, Nano Banana 2 needs to be used to achieve a precise synthesis of a given image by means of a “reference + hint”. This approach is also referred to as the “reference drawings”。
The reference diagrams are indicative, the basic skeletons are the same as the standard texture formula in the front. Flexibility:
• SIMPLICITY: THE SUBJECT DESCRIPTION OF A SINGLE SENTENCE CAN BE WRITTEN, SUCH AS “SMOKING THE MAN IN FIGURE 1 IN THE SCENE IN FIGURE 2”. BUT IT'S EXTREMELY RANDOM, AND IT'S BASED ENTIRELY ON AI'S PREFERENCES。
• STANDARDIZED WRITING (RECOMMENDED): CONTINUE TO APPLY OUR GOLD FORMULA (STYRE)+ (SITUATION)+ (SUBJECT DESCRIPTION)+ (ENVIRONMENT)+ (PHOTO)+ (MASS). IT IS ONLY NECESSARY TO SPECIFY “WHO (FIGURE X) IS IN WHAT SCENE (FIGURE Y) AND WHAT IS IN IT (FIGURE Z)” IN THE CONTEXT OF BOTH “SUBJECT DESCRIPTION” AND “ENVIRONMENTAL SCENE”。
1) Case I: Integration of roles and scenes
Figure 1
Figure 2
- Large film-class screenshots, medium view lens (over and above waist) and side shot. Please place the male hero in figure 1 seamlessly in the late night-over-sea bridge scene in figure 2, where he is tired of smoking on the edge of the bridge. Neon is in a cold tone, and the pink blue light on the bridge is naturally reflected on the side of the master's face and on the windshirt。
Can you see the difference between the effects of different models
2) Case II: Production of the cover of the short play
GPT Image 2 has a great advantage in making cover. It has a strong ability to render words, with simple instructions, without complex hints, to generate very well-designed text, leading to very good images of covers, posters, brochures, instructions, etc. GPT Image 2 is amazing, and it's called a miracle。
The cover of an ancient short play called "The White-Eyed Wolves of Rebirth" is gone, and the leading actress in figure 1 is in the center. There were only four of them。
Snippet script template
If it is hoped that it will be possible to do the film in accordance with the traditional hand spectroscopy (disassembly all scripts into one spectroscopy, each spectroscopy will be produced in advance, and the video will be regenerated), the initial spectroscopy can also be generated by means of the following spectroscopy. This is followed by manual optimization。
●
YOU'RE A SENIOR FILM EDITOR, VISUAL DIRECTOR, AND AN AVS DESIGNER. BASED ON THE SCRIPT I HAVE PROVIDED, PLEASE BREAK DOWN INTO A LENS SCRIPT TABLE SUITABLE FOR THE IA VIDEO PRODUCTION AND DIRECTLY GENERATE AN AI DRAWING HINT FOR EACH LENS IN THE LAST COLUMN OF THE TABLE。
"Script content"
♪ Put the script here ♪
## Output Requirements
Output only one Markdown table, do not output additional instructions. The tables are as follows:
Spectroscope, spectro, spectro, spectro, spectro, spectro, spectro, spectro, spectro, speci
## MIRROR RULE
The translation of script literature into filmable camera language and the absence of non-photo-capable descriptions such as “the mind, awareness, sudden understanding, despair” must be translated into action, expression, eyes, micro-expression, physical reaction, environmental change or props。
The spectroscope shall be in line with the commercial short play/drama rhythm, with attention to the lens interface of the “acting reaction” “causal fruit” and the deletion of the meaningless excess。
The scenery is usually propelled by a “middle-sight vision” feature; in case of shock, inversion, conflict, key props, a bipolar lens can be used to jump。
4. If the line for a single role exceeds 30 words, it is to be split into multiple spectrometers to avoid an excessive length of a spectrometer。
The single lens for orgasms, conflicts, fighting plays is contained in 1.5-3 seconds; the single mirror for mattresses, emotions, dialogue plays is controlled in 3.7 seconds; the key reverses can be appropriately extended to 5-8 seconds。
6. The close proximity of the same person to the next lens is subject to the principle of 30 degrees, changing the angle, image or lens movement of the photograph and avoiding duplication。
7. Multiple spectroscopys in the same scene must be consistent with the overall tone, light logic and environment; the scene will be changed before the tone and atmosphere are reset。
## SPECTRUM TIP RULE
1. Each reminder must be in Chinese, and the style must be unified: the reality video style。
2. The hint shall include: style, spectrometry, shot angle, subject, action or pre-action state, environmental background, overall light tone, image atmosphere, 9:16 scale。
3. The graphs should be concise, be quiet and not simply reproduce “image content” and be logically extended in conjunction with the drama logic。
4. If a person appears in a reminder, do not describe the person ' s clothing and appearance and replace it with a role name。
Each reminder must be complete and independent, and no context-dependent references such as “the previous lens, the same scene, the same person, the continuation of the scene”。
The presentation is intended to pave the way for the generation of follow-up videos, so that priority is given to depicting images of " actions before " or " actions are about to occur " , rather than the results after the actions have been completed。
N.B. In the rule part of the hint, the style and scale of the film to be produced for yourself needs to be adjusted manually。
(iv) Optimization of the lens and image remediation
Picture fast fission
Repeated redrawings on the same map can result in severe coded wear and tear, leading to a decrease in the sharpness of the bottom map and the gradual appearance of a smear or noise. To put it bluntly, it's just that it's going to change。
If models are not allowed to be modified, they are allowed to extract their features from existing images and then rereference them on this basis, so that the quality of the images obtained is largely intact. In this way, not only is it possible to avoid the risk of damage to the painting, but it is also possible to create, through a high-quality bottom map, numerous new lenses that meet the requirements of the story。
The specific play is as follows:
1) Multi-perspective conversion
After obtaining a satisfactory single image of the person, use the Nano Banana 2 (Libtv called Lib Navo 2) model and enter: " Generate XX Perspectives/XX based on this map"
- Generate a woman's face close-up, and the direction of action and vision remains unchanged。
If the angle of the switch is too different from the original figure, for example, if it turns to a woman looking out of the window, then it is clear what is out of the window。
- A shoulder lens is generated, the right edge of the outlook is the woman ' s shoulder, the outlook is blurred and the focus is on the window to which the woman looks, the park。
Of course, we can also generate a map of the nine-gauge at different angles in a batch of Nano Banana 2/GPT Image 2 and then extract the desired figure。
- According to this map, women's positions and movements must remain the same, with a ratio of 16:9 in the 3*39 palaces, with clear numerical serial numbers on the upper left corner。
- Draw the 2nd graph, then zoom in and remove the digital watermark from the upper left corner. 16:9。
2) LibTV Multiple angles
The angle of the picture can also be adjusted by a key to adjust the nodes by the angles embedded in LibTV, Loveart and Tapnow。
Upload one to adjust angle
Select Multi-angle Editor
It includes six predefined angles, including fish eye perspectives, tilt perspectives, head-overs, front-up, panorama, back-up perspectives, and, of course, self-defined angles. A map only needs a point. It's good。
3) Key to precision control: role position maps
For relatively complex indoor scenes, it would be useful to have a positional relationship map of the person in advance as a reference。
Like this panorama, space relations are clear, and there is no need to upload other scenarios。
Role Location Diagram
How exactly do you produce site maps like the following
It's going to be a lot of content, and it's going to be more specific in the field of the fifth issue, and it's going to be based on our article, in conjunction with the 3D channel in LibTV: I built a 3D set with LibTV, which is a perfect solution to the problem of AI video multi-persons and multiple perspectives
4) Video frame modifications
When the long video material is partially deformed, the last perfect image before the collapse can be extracted from the editing software。
Use GPT image 2/Nano Banana to create the next consistent action image with a hint. It is then used as a frame reference to regenerated the second half of the video in the video model, and finally to collage in the editing software。
Through Nano Banana 2/ GPT Image 2, enter a hint: "Take off objects held by a man's hand, put your finger forward and the rest remain unchanged."
Extract flawed maps
Optimized Figure
For this short video material, the actions and mirrors are not complex and can be used without the Feedance 2.0 model, with the initial frame function of the kling o3 model, and with the same support for sound drawings, which is fully effective。
The minimum generation time for a dream-based Feedance 2.0 model is 4 seconds, while the minimum generation time for a vision 1.5 pro model is 5 seconds. Because it's a simple move, it's just a connection between two fixed lenses, and the lines are extremely short, and usually two to three seconds is enough。
Vidu Q3 is the shortest time possible to produce 3 seconds of content for the Vidu video 3.0 Omni (kling o3) model, and the shortest time to produce 1 second, so Vidu or Vidu may also be used in time-appropriate terms。
Sound design
Current mainstream AI video models, such as feedance 2.0, Vidu 3.0, Wan 2.7, HappyHorse 1.0, etc., all support the use of sound drawings, i.e., we can direct the image + music + sound + voice。
THIS DIRECT EFFECT MUST HAVE BEEN VERY APPROPRIATE, WITH THE MIXER TECHNIQUE MORE SPECIALIZED THAN MOST ORDINARY PEOPLE, AND THE MIXER PIECE, WHICH IS ALSO A FULL-OF-EMOTIONAL, TIGHT-WIRE, SAVES US FROM THE LATE-TO-MOUTH-STYLE PROCESS, WHICH HAS GREATLY REDUCED THE AI SENSE FOR THE AMI SHORTS。
BUT IF YOU WANT TO DO A FULL SET OF AMI/AI DRAMAS, YOU CAN'T DO IT WITH VIDEO MODELS. IN MANY CASES, WE STILL NEED TO BE SUPPLEMENTED BY TRADITIONAL LATER MUSIC, SOUND, AND VOICE。
(i) Music
MUSIC APPLICATIONS IN THE AI SHORT PLAY
THE SPECIAL FEATURES OF THE SHORT PLAY: THE TRADITIONAL REALITY SHOW, 40 MINUTES, HAS A COMPLETE OST STRUCTURE. MANY OF THE AI SHORT DRAMAS ARE USUALLY SHORT-SCREENED VIDEO LOGIC, WITH A SINGLE SET LASTING ONE TO THREE MINUTES AND A VERY FAST PACE. SO THEIR MUSIC USUALLY HAS THE FOLLOWING CHARACTERISTICS:
• GO TO OP/ED: THE SHORT PLAY IS RARELY FOLLOWED BY A LONG STORY, AND USUALLY GOES DIRECTLY INTO THE DRAMA。
• FUNCTIONAL BGM: SHORT PLAY MUSIC IS MORE FOCUSED ON “FUNCTIONAL”, I.E., RAPIDLY REKINDLING AUDIENCE EMOTIONS IN A SHORT TIME。
short dramas are limited in size and traditional headleafs are usually omitted and only functional bgms are retained, mainly for emotional use。
Where are we going to get these music files
THERE ARE TWO MAIN WAYS TO GET BGM: ONE IS CREATED WITH AI AND THE OTHER IS TO GO TO THE MATERIAL LIBRARY AND FIND IT. BOTH APPROACHES HAVE ADVANTAGES AND DISADVANTAGES。
1) GENERATE ORIGINAL SHORT PLAY USING AI TOOLS BGM
This approach is now the mainstream, especially for those platforms that have a hard sense of copyright (e.g., pipelines), or when you have a very specific picture in your head, but you can't find it online。
ITS ADVANTAGES ARE CLEAR: THE MUSIC THAT IS GENERATED IS UNIQUE, AND THERE'S LITTLE TO WORRY ABOUT SENDING OUT BECAUSE COPYRIGHTS ARE BEING SILENCED, AND AS THE AI MUSIC MODEL GROWS STRONGER, WE CAN PRODUCE ALMOST ALL KINDS OF MUSIC YOU CAN THINK OF。
There are also, of course, disadvantages, it's kind of like a "black-blind box" and it's not so good at certain styles, such as ancient wind, and it's a little high。
SOME OF THE COMMON AAI MUSIC TOOLS ARE SUMMARIZED BELOW:
LIST OF MAINSTREAM AI MUSIC TOOLS AT HOME AND ABROAD
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
We're going to show you how to generate music with Suno. So we're using Suno as the core presentation tool for this one, and, compared to other tools, Suno is still in a dominant position in the organization of pure musical instruments, in the clarity of the sound quality and in the consistency of long paragraph generation, and is well placed to generate short play with film quality。
here we go into the suno's real business:
Open the main interface for the website
And as you can see, the left is the interface, and the right is the song you created。
It's also very easy to use。
It is divided into the Simple (simple) model and the Advanced (advanced) model。
If you just want a simple tone, Suno's Simple Mode is the fastest entry。
So for the short play, most of the time we just need to generate a pure music BGM that resonates with emotion, and the same is the same, just to open the Instrumental switch and then fill in the description bar with a style hint。
THE FOLLOWING ARE SOME OF THE BGM STYLE TIPS FOR SEVERAL TYPES OF EMOTIONS THAT ARE COMMON IN THE SHORT PLAY/DRAMA, WHICH YOU CAN USE DIRECTLY:
The Simple model is convenient, but it does not control the rhythm. Maybe it just needs a mattress to start with, and it comes up and gives you a climax。
If you want precise control and if you want to customize lyrics (if necessary), select the Advanced model。
If this episode is relatively long and the subject matter is the same, we need the music itself to have some emotional ups and downs, not to come to an orgasm。
This is when we need to come up with specific instructions for different paragraphs of music, called Metatags。
The meta-label is a command packaged in square brackets. The Suno model identifies the contents in square brackets as " structural control instructions " rather than " voice generation instructions " , so as to implement the music instructions without performing a performance。
Switches to the Advanced mode, selects the model and enters the lyrics/commands with a meta-label in the Lyrics bar, and then enters the arc in the Styles bar below. This will be done as follows:
USUALLY, COMPLETE POP MUSIC SONGS HAVE THE FOLLOWING FOUR TO FIVE PARAGRAPHS, WHILE THE COMPLETE SHORT PLAY BGM HAS THE SAME STRUCTURE。
WE CAN EASILY GENERATE WORDS WITH META-LABELS THAT MEET THIS DEMAND THROUGH AI. IN THE CLASSROOM, YOU WILL ALSO BE TAUGHT HOW TO QUICKLY GENERATE, THROUGH SIMPLE TEMPLATES, A PURE, EMOTIONAL, SHORT-TIME MUSIC BGM WITH A COMPLETE STRUCTURE。
2) Clip music material library (regarding copyright)
IN ADDITION TO USING THE AAI MUSIC TOOL TO GENERATE AAI MUSIC, WE CAN FIND SOME OF THE AVAILABLE MUSIC MATERIAL DIRECTLY。
The most common approach is to search for music directly in the music library of the clippings。
find the music bank, and you can be classified, and you can search directly。
Note: In order to avoid copyright problems, it is advisable to choose commercial material。
ii) Sound
The sound effect is not just a reduction of the sound, but a rhythm regulator。
We usually have three layers in our production: a sound that creates a sense of space, a sound that increases a sense of truth, and a special sound that emphasizes visual shocks。
How to Get Sound
There are currently two main ways to access sound effects:
THE FIRST IS THE SOUND GENERATION FUNCTION USING THE AI TOOL. NOW THE BIG MODEL IS READY FOR THE SAME SOUND, NOT TOO MUCH LATER。
IF YOU WANT TO DO IT ALONE, YOU CAN DO IT. MANY OF THE AAI VIDEO TOOLS NOW HAVE AUDIO-GENERATED FUNCTIONS, E.G., CURING, I.E., DREAMING, FILMING ME。
The methods used are very different
When you enter the interface, switch to sound generation, where you can select the voice and video。
Select the textual sound, enter a hint, and you need any sound to generate。
The audio effect of the video is to allow you to upload a video material directly, which automatically produces the sound you want, and below which you can enter a hint to describe it or not. This automatic generation, although limited in precision, is convenient。
IN ADDITION, AN AI SOUND CAN BE GENERATED IN A CUT。
ALL YOU HAVE TO DO IS SELECT AN AI SOUND FOR ANY ONE OF THE MATERIAL, RIGHT-CLICK. IT'S CERTAINLY NOT AS NATURAL AS SOUND MATERIAL, WHICH IS GENERALLY RECORDED. HOWEVER, THE SOUND GENERATED BY AI IS SUFFICIENT FOR SOME SIMPLE ACTIONS, SUCH AS FOOTSTEPS, OR FOR SOME MORE ABSTRACT IMAGES。
The second is the traditional stacking of materials, which is also a more professional method。
The most common is to search for the corresponding sound effects directly in the sound libraries of the clippings, which almost meets the sound needs of most of your regular videos。
I also recommend some resource sound downloads: love for the Internet
(iii) Sounding (not necessary)
Since the advent of models such as Seedance 2.0, all lines can theoretically be generated in tandem with video material, provided that sound assets are collected in advance. This has already been discussed in section III and will not be further elaborated here. The next presentation of the AI voice tool, which is not currently necessary, is self-employable if interested。
USUAL AI SOUNDING TOOL
The following main tools have been screened for different production needs:
DubbingX: The rich language and emotional control that makes it possible to fine-tune whether the word is "aggressive" or "magic" and is well suited to fine-tuning。
MiniMax: Its core strength is extreme accuracy. It generates speech atrophs, breath and natural stasis, which sound very human and suitable for large-scale natural dialogue. It is also particularly for English. The overseas version has voice cloning functions and is very useful。
The dream: the dream is a picture/video and a voice, and its cloned sound function is equally useful in certain situations requiring emergency response。
ElevenLabs: an internationally recognized pole suitable for multilingual or overseas content。
IndexTTS2 / GPT-SoVITS/Qwen-TTS: Local Open Source Tool. For students with a technical background, the sound model of a given role can be trained。
WELL, THAT'S THE CORE PART OF MY WHOLE AI SHORTS。