existAI PaintingIN THE CIRCLE, I FOUND A VERY COMMON PHENOMENON: WHEN YOU SEE AN AMAZING PICTURE, THE FIRST REACTION IS OFTEN -- "DROW IT TO AI AND LET IT TELL ME WHAT THE HINT IS."
Whether it uses Midjourney ' s /describe or throws pictures to CHT GPT to describe them, many people think that if they get the text that AI returns, they can recreate the original image. But reality is often cruel: the hint you get, the picture that comes out of it, but the key tastes are gone。
IN ESSENCE, THIS IS NO DIFFERENT FROM THE “TICK CARD” IN THE GAME. WHY? BECAUSE YOU IGNORED THE CORE CONTRADICTION OF THE LOGIC OF AI GENERATION: THE “OUTCOME” THAT AI SEES IS OFTEN NOT THE “CAUSE” OF THIS PICTURE。
WHEN AI TURNS THE PICTURE, IT'S DOING "READ UNDERSTANDING," IT'S DESCRIBING "WHAT'S IN THE PICTURE," BUT OUR HINT IS "PAINTING INSTRUCTIONS" AND IT'S TELLING AI "HOW TO DRAW." THERE IS A HUGE GAP BETWEEN THE TWO. AI IS EASILY ATTRACTED TO THE DETAILS IN THE PICTURE (E.G., THE RAISINS OF CLOTHES, THE ACTIONS OF ROADERS), THUS IGNORING THE UNDERLYING KEY CONDITIONS THAT REALLY DETERMINE THE IMPACT OF THE IMAGE。
IF YOU SIMPLY RELY ON AI'S AUTOMATIC DESCRIPTION, YOU'LL ALWAYS GET A VERY RANDOM WORD。
TODAY, THROUGH THREE CORE METHODS, WE HAVE THOROUGHLY CORRECTED THE WRONG LOGIC OF THE INVERSE HINT AND TAUGHT YOU HOW TO USE AI TO TRULY “STEAL” THE SOUL OF A GOOD PICTURE。
Method I: Lock the constant entry
1.1 WHY CAN'T AI READ YOUR HEART
When you see a picture, your brain actually has a round of screening. You think this is "good" because you're hit by a particular hit。
Maybe it's the Neon color of Sabourunk
Maybe it's the feeling of loneliness in the eyes of the person
Maybe it's the kind of oppression from the bottom up。
BUT WHEN YOU THROW IT TO AI, AI DOESN'T KNOW WHAT HIT YOU. IT'S LIKE, "THERE'S A PICTURE OF A MAN IN BLACK WITH A BUILDING IN THE BACKGROUND..."
THIS LEADS TO A MISDIRECTION OF INFORMATION: WHAT YOU WANT IS FLOODED WITH AI。
1.2 Operational core: Think first and ask questions later
BEFORE REQUESTING AI'S BACKSLIDING, YOU MUST COMPLETE ONE STEP OF MANUAL INTERVENTION — LOCKING IN THE “CHANGE ITEM” IN YOUR HEART。
The so-called "no change" item is the one that if you remove this element, this figure loses its reference value to you。
1.3 Practising presentations
Assuming you have a very nice view, your favorite is its image and color。
Wrong command:
"Please describe this picture and give me a hint."
Result: Perhaps the image can be restored, yet the image will not be restricted, and the location of the person and the moon will be very random。
Correct command (lock unchanged):
"I need to reverse the hint of this picture. Please note that the most attractive part of this picture is my image and color. As a professional photographer, you are requested to focus on analysing and extracting key words that recapitulate this image and colour effect and to retain the core without change items.”
1.4 Methodological summary
All you have to do is ask questions with answers. Don't ask AI, “What is this”, but ask AI, “How should I write if I want to keep this feature”. In this way, the weight of the hint is naturally tilted on the elements you care most about, thereby ensuring the stability of the result。
Method II: Generation of structural skeleton
2.1 Traps of style words
Many beginners are very obsessed with collages of "styre words" when inverting the hints. Cinematic, Dreamy, Masterpiece, 8k resolution..
Do those words work? It works, but they're just decorations. Like building a house, these words are wallpapers and carpets. But if you don't have pillars and beams, the house can't be built。
It's not style, it's structure。
2.2 What is a “structured skeleton”
The hard core conditions that really cut across the entire AI generation process usually consist of three dimensions:
Image: Where are the cameras? Do they look down, look up or look down? What's the focus
Lighting Logic: Where did the light come from, side-reverse light, top light or flashing light
Space Relations: Where is the subject? How does the distance between the background and the subject feel
IF YOU GIVE ONLY STYLE WORDS, AI WILL START TO BUMP RANDOMLY ON THE "CONSTRUCTION" AND "LIGHT." THAT'S WHY THE WORDS YOU'RE INTRODUCING ARE SOMETIMES BIG HEAD PICTURES, SOMETIMES ALL OVER YOUR BODY, SOMETIMES EVEN LANDSCAPES。
2.3 Practical demonstration: how to extract the skeleton
We need to force AI to decipher the image from the perspective of the Generation System, not from the view of Viewer。
Wrong command:
"This is a very cinematic picture. Please write a hint."
Correct command (extract structure):
"Please analyze the image. Do not use adjective stacking. Please reverse the hint by the following three dimensions:
Camera space and focus; light layout; geometrical map. Please give me a set of hints to focus on these structural skeletons."
2.4 Summary of methodological discussions
THE STYLE IS SKIN, THE STRUCTURE IS BONE. WHEN USING AI, IT MUST BE FORCED TO ANCHOR THE SKELETON OF THE IMAGE THROUGH “TECHNICAL TERMS”. WHEN THE STRUCTURE IS FIXED, YOU FILL IT WITH ANY STYLE (E.G., CHANGING SABERPUNK TO STEAM PUNK), THE PICTURE REMAINS VALID AND VERY STABLE。
Method III: Inverse error point of validation
3.1 Fatal injuries of a newcomer: one-time delivery
THE EASIEST MISTAKE FOR A NEWCOMER IS “ONE-OFF TRUST”. TAKE THE WORD AI'S INVERTED, GO STRAIGHT TO THE RUN, FIND IT'S NOT LIKE IT, AND THEN FEEL "AI CAN'T" OR "NO" AND THEN GIVE IT UP。
It's a huge mistake. The inverse reminder is not an " action " , but a " process " 。
IT'S ALSO AN ILLUSION. IT MAY MISTAKENLY IDENTIFY A SHADOW IN THE PICTURE AS A BLACK CAT OR A TEXTUALIZED BACKGROUND. IF YOU DON'T VERIFY, THESE MISTAKES WILL BE BROUGHT INTO YOUR CREATION。
3.2 Introduction of “correction mechanisms”
WE NEED TO SET UP A CLOSED LOOP: BACK-UP-GENERATION-COMPARISON-CORRECTION. THE MOST CRITICAL STEP IS TO GET AI TO DO IT HIMSELF。
3.3 Practised demonstration: three-step validation
STEP 1: PRELIMINARY INVERSION AND GENERATION OF THE A-INDICATOR THAT ALLOWS AI TO REVERSE IMAGE A, AND THEN USE THIS TO GENERATE PICTURE B。
Step 2: Visual Comparison sends both picture A (original) and picture B (new) to AI (beans bag or other visual model)。
Step three: Self-correction instructions
"PLEASE COMPARE PICTURES A (ORIGINAL) WITH PICTURES B (I.E., I PRODUCE WITH THE HINT YOU PROVIDE)。
YOU'LL FIND THAT THERE'S A CLEAR DIFFERENCE BETWEEN IMAGE B ON THE MASS/CONCEPT/PHOTO/PHOTO. PLEASE ANALYSE: WHAT ARE THE KEY TERMS OF THE MESSAGE MISSING IN PHOTO B THAT LEAD TO THIS DIFFERENCE? IS THE LIGHT GOING IN THE WRONG DIRECTION? IS THE MATERIAL DESCRIPTION NOT ACCURATE
ON THE BASIS OF THIS VARIANCE ANALYSIS, PLEASE MODIFY AND OPTIMIZE THE PRE-INDICTMENT SO THAT THE IMAGE PRODUCED CAN BE MORE ACCURATELY REDUCED TO THE EFFECT OF IMAGE A.”
3.4 Methodological summary
In this way, you used AI's analytical capabilities to fix its own loopholes. You'd be surprised to find that after a round of "correcting" phrases, there's often more "magic" words than you'd ever thought。
This will not only help you to get a precise hint, but it will also help you to find the conditions under which it really works — that is, “which is the word that determines the success or failure of this map”。
Appendix: Summary of all-powerful inverse formula
In order for you to get your hands on it, I condensed these three methods into a ** "mathematical formula" and a set of "one-size-fits-all command templates"**. Then there'll be a map that you want to reset. Just apply it。
1. Bottom logical formulae
We split a perfect hint into three variables:
PERFECT HINT = A (CORE NO CHANGE) + B (STRUCTURED SKELETON) + C (STYLE MODIFICATION)
NOTE: TRADITIONAL ERROR INVERSION TENDS TO FOCUS ONLY ON (STYLE) AND IGNORES DECISIVE A AND B。
2. AI INTERACTIVE COMMAND TEMPLATE (DIRECT REPRODUCTION)
DO NOT SEND ONLY ONE MAP TO AI. PLEASE SEND IT ALONG WITH THE FOLLOWING COMMAND SPELL. A HIGH-QUALITY INVERTED RESULT CAN BE OBTAINED BY SIMPLY FILLING IN BLANK BRACKETS。
Step 1: Initial inversion directive
Please act as a professional photographer/shower/film director to analyze the uploaded picture。
My goal is to reverse the hint of this picturePromptAnd it recreates a very similar image。
Please proceed with dismantling in strict accordance with the following logic and not with excessive literary dispersion:
1. [LOCK UNCHANGED ITEM A]:
This is the most attractive image of me. Please ensure that the maximum weighting of this part is given in the reminder。
2. [EXTRACT STRUCTURAL SKELETON B]:
Please use professional terminology to analyse the physical composition of images:
– Constructive approach (e.g. trio, centre, corner Netherlands...)
– lens language (e.g. 35 mm focal, shallow view, perspective...)
– Light layout (e.g. Rembrandt, Dindal effect, volume light...)
3. [SYSTEM C]:
The last of these is style genres and rendering details (e.g. Fantasy Engine Five Rendering, Gypsy...)。
On the basis of the above analysis, please output a clearly structured English phrase (Prompt)。
Step 2: Overcoming the trajectories (if the first step does not produce a picture)
I USED THE HINT YOU JUST GAVE ME TO CREATE A PICTURE OF B, BUT IT FELT LIKE IT COMPARED TO THE ORIGINAL A, WHICH WAS THE FIRST。
Please compare the two graphs
1. What are the key [structural conditions] or [photographic description] missing in the message
2. Please modify and optimize the previous version of the introduction and focus on the amendment [insert questions for amendment]。
Concluding remarks: From Descriptor to Builder
THE CORRECT USE OF AI'S INVERTED HINT IS CERTAINLY NOT FOR AI TO ACT AS A READER THAT CAN ONLY “READ AND SPEAK”。
YOU NEED TO THINK OF AI AS A SYSTEM ENGINEER WHO CAN TALK。
Lock the constant: tell it what the core soul is。
Generating structural skeletons: Telling it to build the physical foundation of the image。
Validation inverted error: By comparing differences, force it to hand over more precise control codes。
When you master these three methods, you are no longer in the "scratch card" but in the exact reverse engineering. This is the high-level game of Prompt Engineering。
Now, take out the picture in your album that's been in your photo album for a long time, and try it again, using today's method。