HOJOON KWEON - Generative AI

PILLAR #3: GENERATIVE AI FUNDAMENTALS

---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

TABLE OF CONTENTS

---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

COURSE 1: BASICS OF QUANTUM INFORMATION [____________________]

⇒ ______________________ [______________________] ⇐

COURSE 2: BASICS OF QUANTUM INFORMATION [____________________]

⇒ ______________________ [______________________] ⇐

COURSE 3: BASICS OF QUANTUM INFORMATION [____________________]

⇒ ______________________ [______________________] ⇐

COURSE 4: BASICS OF QUANTUM INFORMATION [____________________]

⇒ ______________________ [______________________] ⇐

COURSE 5: BASICS OF QUANTUM INFORMATION [____________________]

⇒ ______________________ [______________________] ⇐

---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

COURSE 1: GENERATIVE AI - INTRODUCTION AND APPLICATIONS [______________________________]

---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

TABLE OF CONTENTS

TITLE

⇒ _____________________: _________________ | _________________ | _________________ | _________________ | _________________ ⇐

TITLE

⇒ _____________________: _________________ | _________________ | _________________ | _________________ | _________________ ⇐

TITLE

⇒ _____________________: _________________ | _________________ | _________________ | _________________ | _________________ ⇐

TITLE

⇒ _____________________: _________________ | _________________ | _________________ | _________________ | _________________ ⇐

GENERATIVE AI AND ITS CAPABILITIES [__________________]

Introduction to Generative AI

⇒ Artificial Intelligence models uses deep learning and neural networks to train Artificial Neural Networks on massive existing data. ⇐

⇒ Discriminative AI: Distinguishes between different classes of data to differentiate, classity, identify patterns, and draw conclusions. [Ex: Email Spam Filters] ⇐

⇒ Generative AI: Creates new content based on training data by putting in the prompt and receiving output which heighten creativity. [Ex: ChatGPT and Gemini] ⇐

⇒ Large Language Models [LLMs] is a foundation model which intends to process and generate text. Open AI created GPT n-series starting from 2018 as their LLMs. ⇐

⇒ The Predicted Potential of Generative AI Market: The Generative AI Market is estimated to reach $1.3 Trillion by 2032 while currently at $20.9 Billion in 2024. ⇐

History and Evolution of Generative AI

⇒ During the 1980s and 1990s, Neural Networks was created with an inspiration from human brain which can learn intricate patterns in data.⇐

⇒ In the early 2000s, Deep Learning was found which utilized neural networks with multiple layers, training extensive datasets to discern complex patterns. ⇐

⇒ The development of Deep Learning led into the Generative AI models including Generative Adversarial Networks [GANs] and Variational Autoencoders [VAEs]. ⇐

⇒ GANs operate by training two neural networks in opposition: Generator which creates new content and Discriminator which tries to differentiate synthetic content. ⇐

⇒ VAEs generates contents by learning and sampling a Latent Space, a representation of the data that captures the most essential features of the data they are trained on. ⇐

Capabilities of Generative AI

⇒ Text Generation: Large Language Models [LLMs] are trained by learning patterns and structures from large datasets and generate coherent and contextually relevant text. The capabilities of LLMs include text completion, summarization, question answering, translation, code generation, image and text pairing, and conversational interaction. ⇐

⇒ Image Generation: Deep Learning techniques including GANs and VAEs are leveraged in the models to create realistic textures, natural colors, and fine-grained details. ⇐

⇒ Audio Generation: The Text-to-Speech Audio [TTS] creates highly realistic speech including details in tone, pitch, modulation, pronunciation, rhythm, and expressions.⇐

⇒ Video Generation: Textual-Prompt based video can specify desired content and guide video generation process to transform image into dynamic video and add motions. ⇐

⇒ Code Generation: LLMs create code snippets, functions, and complete programs. In detail, the features include synthesizing code [creating based on given requirements], refactoring code [restructuring without changing functionality], fixing bugs in the code, testing software, and creating documentation including comments and usages. ⇐

⇒ Data Generation and Augmentation: Generate new data and augment existing datasets [generating new data from existing data] leading to higher diversity and variability of data. Specifically, new samples and augment datasets can be generated including texts, images, speech, tabular data, statistical distribution, time series data, and finance. ⇐

⇒ Virtual World Creation: Generate virtual avatar simulations [realistic behavior, expression, and conversation] and complex virtual environments [realistic experience]. ⇐

APPLICATIONS AND TOOLS OF GENERATIVE AI [__________________]

Applications of Generative AI

⇒ IT and DevOps: Improves software delivery processes and infrastructure management by reducing manual coding efforts and time spent on repetitive tasks. The models generate synthetic test cases by simulating user behavior and impacting software efficiency, detect anomalies by analyzing system logs and metrics, and automate process. ⇐

⇒ Entertainment: Generates synthetic contents such as music, scripts, stories, videos, movies, video games, and utilized for translation, localization, and personalization. ⇐

⇒ Education: Generates personalized learning experience and simulated experiential learning. Also, the algorithms detect special needs and create flexible lesson plans. ⇐

⇒ Banking and Finance: The algorithm auto-detects risks, generates insight, provides financially literate recommendation, and simulates potential fraud scenarios and credit scoring. Moreover, the model conducts market sentiment analysis [automatically analyzing news, articles, social media, and classified textual data] and algorithmic trading. ⇐

⇒ Healthcare: The algorithm conducts Medical Image Analysis [synthesizing images resembling patient data and rare medical conditions data], Drug Discovery [generating molecules and speeding up drug discovery processes], and Telemedicine, or Telemonitoring [establishing immediate medical advice and personalized medical treatment] . ⇐

⇒ Human Resources: Automates repetitive Human Resources Tasks such as creating job requisitions, screening relevant profiles, scheduling interviews, and onboarding. ⇐

⇒ McKinsey Report on Economic Potential of Generative AI: Current GenAI possess the potential to automate work that absorbs 60% to 70% of employee's time today. ⇐

Tools of Generative AI

⇒ Text Generation: LLMs are based on patterns learned on training to interpret contexts, grammars, and semantics [draw statistical relationship between words and phrases to adapt creative writing style]. The models include GPT [effective in dynamic response and conversational flow] and PaLM [optimal to research current topic accurately]. ⇐

⇒ Image Generation: Common Image-Generation models include DALL-E, Stable Diffusion, and StyleGAN with image generation, transformation, and API integration. ⇐

---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

COURSE 2: GENERATIVE AI - PROMPT ENGINEERING BASICS [______________________________]

---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

TABLE OF CONTENTS

TITLE

⇒ _____________________: _________________ | _________________ | _________________ | _________________ | _________________ ⇐

TITLE

⇒ _____________________: _________________ | _________________ | _________________ | _________________ | _________________ ⇐

TITLE

⇒ _____________________: _________________ | _________________ | _________________ | _________________ | _________________ ⇐

TITLE

⇒ _____________________: _________________ | _________________ | _________________ | _________________ | _________________ ⇐

CONCEPT OF PROMPT ENGINEERING [__________________]

Definition and Elements of a Prompt

⇒ Quotation by a French Anthropologist Claude Lévi-Strauss: "An expert knows all the answers if you ask the right questions." ⇐

⇒ Prompt is the critical factor of generating the output which are relevant, contextual, imaginative, nuanced, and accurate which closely resembles the output of humans. ⇐

⇒ Prompt: An input that produces the desired input, which can also be a series of instructions that refine the output step by step. ⇐

⇒ Ex: "Write HTML code to generate a dropdown selection of cities within an online form." "Write a short story about a scientist studying life on Mars." ⇐

⇒ Based on the natural language requests submitted as prompts, models collect information, derive inference, and provide creative solution with relevance and logic. ⇐

⇒ To convey correct intentions to the models which are not brief and lacking details, make adjustments by including context, proper structure, and comprehensibility. ⇐

Building Blocks of a Well-Constructed Prompt

⇒ The following four elements play roles in helping the Generative AI models to understand the requirements and provide desired output. ⇐

⇒1) Instructions: Provide distinct guidelines regarding the task you wish to execute, for example, "write an essay in 600 words analyzing the effect of global warming." ⇐

⇒ 2) Context: Establish settings and framework for generating relevant content by providing the background information about the overarching topic within the prompt. ⇐

⇒ 3) Input Data: Provide any pieces of information which help the model to attain responses with a specific set of details, for example, a dataset of temperature records. ⇐

⇒ 4) Output Indicator: Offer benchmarks and criteria for assessing attributes of the output generated including tone, style, length, clarity, or any other desired qualities. ⇐

⇒ Output Indicator Ex: "The output should be 600 words. It will be evaluated based on the clarity of analysis and incorporation of relevant data or case studies." ⇐

Definition of Prompt Engineering

⇒ Prompt Engineering: Process of designing effective prompts to generate better and desired responses. ⇐

⇒ The failure to provide precise prompts may lead the models to produce inadequate results, and more critically, false and misleading information. ⇐

⇒ Prompt Engineering is a blend of critical analysis, creativity, and technical acumen, weighing in framing the question to elicit the most appropriate response. ⇐

⇒ Ex: When a ship is sailing through the Atlantic Ocean, a captain plans his voyage and wants to know the precise weather forecast for a specific location at a specific time. Asking "Weather forecast of the Atlantic Sea" is a right question, but may not provide the desired result because details on context are not defined including intended location, longitude, and time range. Instead, the prompt should state "The captain of a ship is planning a strategic voyage in the Atlantic Ocean. To help the captain navigate effectively, provide weather forecasts for the upcoming week from 28th August 2023 to 1st September 2023. The coordinates of the target location are between 20 Degrees North and 30 Degrees North Latitude and 40 Degrees West and 20 Degrees West Longitude. To help plan an effective navigation in the Atlantic Ocean, provide detailed information about expected wind patterns, wave heights, precipitation probabilities, cloud covers, and potential storms that might affect the voyage with a specified time frame and location." ⇐

Processes of Prompt Engineering

⇒1) Define the Goal: Establish a distinct goal with an understanding of what you exactly want the model to generate. ⇐

⇒ Step 1 Ex: "Form a brief overview of the benefits and risks associated with artificial intelligence in automobiles." ⇐

⇒2) Craft Initial Prompt: Depending on a defined goal, create an initial prompt which can manifest as a question, a directive, or a situation.. ⇐

⇒ Step 2 Ex: "Write an article that presents an analysis of the benefits and drawbacks associated with the incorporation of AI in the field of automobile industry." ⇐

⇒ 3) Test the Prompt: The response created from the prompt must be tested since the response might be relevant but lack an aimed distinct perspective. ⇐

⇒ Step 3 Ex: The output to the initial prompt directly lays out the benefits and drawbacks of integrating AI, but not any of the ethical concerns and implications. ⇐

⇒ 4) Analyze the Response: Carefully review the response and examine whether it aligns with the goal, and if not, make a note of the areas where it fell short. ⇐

⇒ Step 4 Ex: The initial prompt failed to cover the comprehensive range of benefits and risks associated with artificial intelligence in the automobile industry. ⇐

⇒5) Refine the Prompt: Using the knowledge acquired through tests and analysis, modify the prompt by enhancing specificity, incorporating context, or rephrasing. ⇐

⇒ Step 5 Ex: The redefined prompt made by testing the initial prompt should read, "Write an informative article discussing the role of artificial intelligence in revolutionizing the automobile industry. Address key aspects including the benefits, drawbacks, ethical considerations, and both positive and negative implications. Cover specific domains including autonomous driving and real-time traffic analysis, while also examining the potential challenges including technical complexity and cybersecurity concerns." ⇐

⇒ 6) Iterate the Process: The last three steps [test the prompt - analyze the response - refine the prompt] are repeated until you are satisfied with the response. ⇐

⇒ Step 6 Ex: Final prompt should read, "Write an article highlighting how artificial intelligence is reshaping the automobile industry focusing on the positive advancements, particularly in autonomous driving and real-time traffic analysis, while thoroughly exploring concerns about intricate technical aspects including decision-making algorithms and potential cybersecurity breaches. Emphasize the implications these concerns may affect on vehicle safety. Ensure that analysis is thorough and backed with examples. ⇐

Importance of Prompt Engineering

⇒ 1) Optimizing Model Efficiency: Help to design intelligent prompts which allow the users to harness full capabilities of the models without extensive retraining. ⇐

⇒ 2) Boosting Performance for Specific Tasks: Empowers models to deliver nuanced and in context responses which let them be more effective for specific tasks. ⇐

⇒ 3) Understanding Model Constraints: Studying the corresponding responses by refining prompts helps to understand the strengths and weaknesses of the model. ⇐

⇒ 4) Enhancing Model Security: Presents issues of harmful content generation which can happen by poorly designed prompts and hence enhancing safe utilization. ⇐

Best Practices for Prompt Creation

⇒ User can apply the best practices across four essential dimensions [clarity, context, precision, and persona pattern] which allows supervision of style, tone, and content. ⇐

⇒ Clarity: A strategy of utilizing concise language to convey instruction easily, avoiding complex jargons and specialized terminologies which are not understood by a broad range of audience [considering the field literature], and providing explicit instructions with a clear description of the task which align with the intended outcome of a user. ⇐

⇒ Context: A strategy of establishing brief introduction or explanation of the circumstances including details like people, places, events, or concepts to support the model. ⇐

⇒ Precision: A strategy of outlining the request specifically if user is searching for particular kind of response by incorporating examples which direct its thought process. ⇐

⇒ Role-Play/Persona Pattern: A strategy of requesting the model to generate responses with a perspective aligned with a specific character or persona including a historical figure, a fictional character, a specific profession, and a lot more. Additional contextual details enable the model to effectively assume a personality and increase accuracy. ⇐

⇒ Role-Play/Persona Pattern Ex: First example is on a specific person, "Act as a marketing expert Seth Godin to provide a list of 10 article titles to promote a book." Second example is on a profession, "Act as a liberal political expert and as a conservative political expert and provide answers on policies to be implemented to address inflation?" ⇐

Common Prompt Engineering Tools

⇒ Prompt Engineering Tools provide various features to optimize creation of prompts, especially for users who are not proficient with natural language processing [NLP]. ⇐

⇒ The Common Functionalities include prompt suggestion, contextual understanding, iterative refinement, bias mitigation, domain-specific aid, and libraries of prompts. ⇐

⇒ 1) Suggestion for Prompts: Tools offer detailed suggestions for optimization of the prompts based on the given input or the desired output. ⇐

⇒ 2) Contextual Understanding: Tools suggest how to structure prompts with contextual communication to help the model know the user intent. ⇐

⇒ 3) Iterative Refinement: Tools iteratively refine prompt based on the initial responses to find the most effective prompt with accurate response. ⇐

⇒ 4) Bias Mitigation: Tools help to mitigate the risk of potential bias and inappropriateness in the responses by avoiding the misleading prompts. ⇐

⇒ 5) Domain-Specific Aid: Tools help to create prompts which are relevant to some specific domains such as legal, medical, or technical fields. ⇐

⇒ 6) Libraries of Predefined Prompts: Tools offer libraries of predefined prompts of various use cases to be customized for the specific needs. ⇐

⇒ The Prompt Engineering Tools [IBM watsonxs.ai, Spellbook, Dust] offer Edit and Test Prompt Features to optimize the prompt until the user is satisfied. ⇐

TECHNIQUES AND APPROACHES FOR WRITING EFFECTIVE PROMPTS [__________________]

TEXT-TO-TEXT PROMPT TECHNIQUES

⇒ In recent years, the NLP and LLMs experienced a significant advancement. As the size and complexity of LLMs continue to increase, concerns regarding the reliability, security, and potential biases have surfaced. The effective use of text prompts is the only possible and promising solution to these arising concerns surrounding LLMs. ⇐

⇒ The following lists include the 7 techniques for using text prompts effective and improve the reliability of output generated by LLMs. ⇐

⇒ 1) Task Specification: The text prompts should explicitly specify the objective of the LLM to increase accuracy of response. ⇐

⇒ Task Specification Ex: "Translate the English Sentence into Korean and Korean Sentence into English." ⇐

⇒ 2) Contextual Guidance: A technique of using knowledge on which text prompts provide specific instructions to the LLMs. ⇐

⇒ Contextual Guidance Ex: "Write a short paragraph on New York City, highlighting its iconic landmarks." ⇐

⇒ 3) Domain Expertise: The text prompts can use domain-specific terminology in specialized fields for accuracy and precision. ⇐

⇒ Domain Expertise Ex: "Explain the causes, symptoms, and treatments for hypothyroidism, including the latest research and medical guidelines." ⇐

⇒ 4) Bias Mitigation: A technique in which text prompts provide explicit instructions to generate neutral and unbiased responses. ⇐

⇒ Bias Mitigation Ex: Write a 100-word paragraph on leadership traits without favoring any gender and providing equal examples from all genders. ⇐

⇒ 5) Framing: A technique by which the text prompts guide the LLMs to generate responses only within the required boundaries. ⇐

⇒ Framing Ex: "Provide a summary of 100 words of the provided article on climate change, focusing on its primary findings and recommendations."⇐

⇒ 6) Zero-Shot Prompting: LLMs trained on large data can perform tasks Zero-Shot which is a method of generating meaningful responses without prior training. ⇐

⇒ Zero-Shot Prompting Ex: Provide a sentence, "Anita bakes the best cakes in the neighborhood." Then, ask the model, "Select the adjective in this sentence." ⇐

⇒ 7) User Feedback Loop: When user is incapable to receive desired answer in one prompt, one can provide feedback on text prompts and iteratively refine them. ⇐

⇒ User Feedback Loop Ex: Initial prompt says, "Write a poem." Once LLM generates a poem, user asks "Make it more humorous." LLM adjusts until satisfaction. ⇐

⇒ 8) Few-Shot Prompting: Few-Shot enables an in-context learning wherein demonstrations are provided in the prompt to steer the model to better performance. ⇐

⇒ Few-Shot Prompting Ex: "Recommend a summer travel destination well known for beautiful beaches." Afterwards, ask the model, "Recommend a city to explore." ⇐

BENEFITS OF USING TEXT PROMPTS EFFECTIVELY

⇒ 1) Increase Explainability: Improving the degree to which a user can understand and interpret the decision-making process of the model and the reasons behind outputs. ⇐

⇒ 2) Address Ethical Considerations: Addressing ethical considerations by ensuring that the LLMs behavior is consistent with ethical guidelines and legal requirements. ⇐

⇒ 3) Foster User Trust: Understanding how the LLM works since users can see a direct influence of instructions on the LLM behavior and creates meaningful interaction. ⇐

INTERVIEW PATTERN APPROACH

⇒ The Interview Pattern Approach involves designing prompts by simulating a conversation or interacting in the interview style. ⇐

⇒ This approach should provide specific prompt instructions to the model in response to which the model asks necessary follow-up questions from the user. Based on the responses user provides to all follow-up questions, the model will draw information relevant to the task, then process the information to provide an optimized response. ⇐

⇒ Interview Pattern Approach Ex: The situation is when user wants the model to behave like a travel consultant and plan travel itinerary for a vacation. The instruction under Interview Pattern will include information such as, "You will act as a seasoned travel expert. Your objective is to engage in a comprehensive trip-planning session. Begin by asking a series of detailed questions, one at a time, to gather all the essential information required to craft the most tailored travel itinerary based on my specific preferences, interests, and budget." Then, model will ask the relevant questions to optimize its answer, where each question builds upon the previous about creating travel preferences. ⇐

⇒ The Interview Pattern Approach allows a dynamic back-and-forth exchange of information interacting with the model to help clarifying queries and optimizing results. ⇐

CHAIN-OF-THOUGHT APPROACH

⇒ The Chain-of-Thought [CoT] Approach is a prompt-based learning that involves constructing a series of questions to guide the model to generate a desire response. ⇐

⇒ With this approach, user can strengthen the cognitive abilities of the Generative AI models and allow the models to better explain the reasoning processes. ⇐

⇒ This approach should break down a complex task into smaller ones through a sequence of more straightforward prompts. Before posing a question directly, user feeds the model with related questions along with corresponding solutions, helping a model to receive new problem and use same strategy to answer with same line of reasoning. ⇐

⇒ Chain-of-Thought Approach Ex: The question is, "Matthew has 6 eggs. He buys 2 more trays of eggs. Each tray has 12 eggs. How many eggs does he have now?" The user provides a similar question, "Mary has 8 rashishes. She used 5 radishes for a dinner. The next morning, she bought 10 more radishes. How many radishes does she have now?" and the solution, "Mary had 8 radishes. She cooked using 5 of them, so 8 - 5 = 3. Next morning, she bought 10 more, so 3 + 10 = 13. The answer is 13 radishes left. ⇐

⇒ Few-Shot CoT: For a math problem, provide a question, a logical solution to that question, and another question that can be solved by applying the same line of logic. ⇐

⇒ Zero-Shot CoT: Provide a question and simply add "Let's think step by step" to a prompt. Model solves the question by going through step-by-step without examples. ⇐

TREE-OF-THOUGHT APPROACH

⇒ The Tree-of-Thought [ToT] Approach is an innovative technique built to expand the advanced reasoning capabilities of the Chain-of-Thought Approach. ⇐

⇒ ToT involves hierarchically structuring a prompt akin to a tree structure to specify the desired line of reasoning and unlock new solutions of complex problems. ⇐

⇒ This approach should generate multiple line of thought, resembling a decision tree to explore different possibilities and ideas, unlike traditional linear approaches. This technique allows the model to evaluate and pursue multiple paths simultaneously, and each idea branches out to create a tree-like structure of interconnected thoughts. ⇐

⇒ A model assesses all possible routes and assign numerical value according to prediction of outcomes, then eliminates lesser promising line of thoughts until the best. ⇐

⇒ Tree-of-Thought Approach Ex: A situation is when a user wants the model to design recruitment and retention strategies and thereby attract skilled remote employees for an e-commerce business. The prompt instructions will be "Imagine three different experts answering this question. All experts will write down 1 step of their thinking and then share it with the group. Then all experts will go on to the next step, etc. If any expert realizes they're wrong at any point, then they leave." Then, the question will be "Act as a human resource specialist, design a recruitment and retention strategy for an e-commerce business, focusing on attracting and retaining skilled remote employees." Providing prior prompt instruction will help the model to think logically to conduct a step-by-step process and also allow the model to consider intermediate thoughts and build upon. ⇐

⇒ Tree-of-Thought Approach Ex: "Simulate three logical experts collaboratively answering a question. Each one explains their thought process in real-time, considering the prior explanation of others and openly acknowledging mistakes. At each step, whenever possible, each expert refines and builds upon the thoughts of others, acknowledging their contributions. They continue until one definitive answer to the question. For clarity, your entire responses should be in a markdown table format. The question is ..." ⇐

TEXT-TO-IMAGE PROMPTS TECHNIQUES

⇒ An image prompt is a text description of an image to be generated. To increase the impact, users use techniques to improve quality, diversity, and relevance of images. ⇐

⇒ Style Modifiers: Descriptors used to influence the artistic style or visual attributes of images. User can modify visual elements including color, contrast, texture, shape, and size to generate aesthetically appealing image. A prompt may include art styles, historical art periods, photography techniques, art materials, and well-know brand or artist. ⇐

⇒ Quality Boosters: Images with low resolution frequently exhibit blurriness and pixelation. Quality Boosters improve overall fidelity and sharpness of the images. Specific terms include noise reduction, sharpening, color correction, resolution enhancement, high resolution, 2k, 4k, hyper-detailed, sharp focus, complimentary colors, and more. ⇐

⇒ Repetition: Leverages the power of iterative sampling to enhance the image diversity by helping to increase the memorability of the model. This technique is particularly valuable when models are confronted with abstract or ambiguous prompts to which numerous valid interpretations are possible. An example prompt would say "A tiny, tiny, tiny, tiny, tiny, tiny, tiny, tiny, tiny, tiny, tiny, tiny zoxy cabin in the heart of the dense, dense, dense, dense, dense, dense, dense, dense, dense, dense, dense, dense forest." ⇐

⇒ Weighted Terms: Refers to the use of words or phrases that can have a powerful emotional or psychological impact. For example, words including free, limited time offer, and guaranteed, are often used in advertising to elicit a sense of urgency, security, and trust. Similarly, words such as luxury, premium, and exclusive are used to create a sense of exclusivity and sophistication. An example of utilizing weighted terms says, "Craft an image of a cozy living room with a [warm: 10] [crackling: 8] [cold: -7] fireplace." ⇐

⇒ Fix Deformed Generation: This technique is used to modify any deformities or anomalies which may impact the effectiveness of the image. Deformities in an image can include conditions like distortion, pixelation, or other image quality issues which can detract from the visual appeal and clarity of the image. An example would say "Mother Teresa with waving hands [disfigured, deformed hands, distorted hands, distorted fingers, bad anatomy, bad hands]." The good negative words are used to mitigate issues. ⇐