Ai Image Generation: Tools and Techniques

• A. Decision trees
• B. Diffusion models ✓
• C. Linear regression models
• D. K-means clustering

Why: Diffusion models learn to remove noise step by step, allowing them to generate images from random noise. They underpin many modern image generators such as Stable Diffusion and DALL-E 2.

• A. It stores the final image file on the server
• B. It describes the desired image so the model can guide generation toward matching content ✓
• C. It sets the monitor resolution for displaying results
• D. It compresses the output image to reduce file size

Why: The text prompt is a natural language description that conditions the model, guiding it to produce an image that matches the described subject, style, and composition.

• A. Training dataset ✓
• B. Output buffer
• C. Rendering pipeline
• D. Inference cache

Why: A training dataset is the large body of labeled or captioned images the model learns from. The quality and diversity of this data strongly influence what the model can generate.

• A. The visible pixel grid of the final image
• B. A compressed mathematical representation where the model encodes and manipulates image features ✓
• C. The hidden layer of the user's computer memory
• D. A backup copy of previously generated images

Why: Latent space is a lower-dimensional representation in which the model stores abstract features of images. Generation often happens in this space because it is more efficient than working directly with pixels.

• A. Hands and text are always stored in a separate file that is lost
• B. The model has learned imperfect patterns for fine details that are hard to represent consistently from training data ✓
• C. The monitor cannot display hands or letters accurately
• D. The generator intentionally avoids drawing people and words

Why: Fine structures like fingers and written characters have many precise variations, and the model may not learn their exact rules reliably, leading to common artifacts that users often need to correct through prompting or editing.

• A. They generate images by directly predicting every pixel in a single forward pass through a neural network.
• B. They learn to reverse a gradual process of adding noise to images, starting from pure noise and denoising step by step to produce a new image. ✓
• C. They combine many existing photographs into a collage by searching a large database for matching regions.
• D. They draw images by tracing edges and filling colors according to a fixed set of geometric rules.

Why: Diffusion models are trained by corrupting images with noise over many steps (the forward process) and learning to reverse that corruption (the reverse process). At generation time, the model starts from random noise and iteratively denoises it into an image.

• A. The image becomes sharper and gains more detail at each step.
• B. The image is progressively replaced by Gaussian noise until it is essentially indistinguishable from random noise. ✓
• C. The image is converted into a text description that gets longer with each step.
• D. The image is split into smaller tiles that are rearranged randomly.

Why: The forward process adds a small amount of Gaussian noise at each timestep according to a noise schedule. After many steps, the original image information is destroyed and the result approximates pure noise.

• A. To store a complete library of training images that it copies from during generation.
• B. To predict the noise (or the clean image) present at a given timestep so that it can be removed during the reverse process. ✓
• C. To compress the image into a single number representing its overall brightness.
• D. To choose which color palette the final image will use from a predefined list.

Why: The denoising network, often a U-Net, is trained to estimate the noise that was added at each timestep. Subtracting these estimates step by step during sampling recovers an image from noise.

• A. They use a single adversarial game between two networks that is guaranteed to converge quickly.
• B. Their training objective is a simple, stable regression-style loss rather than an adversarial min-max game, which tends to avoid mode collapse and training instability. ✓
• C. They only generate very small images, which reduces the chance of errors.
• D. They skip training entirely and rely on hand-written rules for each pixel.

Why: GANs require balancing a generator and discriminator in an adversarial game that can be unstable and prone to mode collapse. Diffusion models train by predicting noise with a straightforward loss, which is generally more stable and yields good diversity.

• A. The number of colors that can appear in the final image.
• B. The number of denoising iterations used to turn noise into an image, which trades off generation speed against quality. ✓
• C. The file format the image is saved in, such as PNG or JPEG.
• D. The number of people who are allowed to view the generated image.

Why: Sampling steps determine how many reverse denoising passes are run. More steps usually yields higher quality but takes longer, while fewer steps is faster but can reduce detail, which is why samplers and step counts are tuned in practice.

• A. The tool's built-in photo editing filters
• B. The model's demonstrated strength in realistic lighting, textures, and detail ✓
• C. The number of languages the interface supports
• D. The file format the tool exports to by default

Why: Photorealistic mockups depend on how well the underlying model renders lighting, materials, and fine detail. Interface languages, export defaults, and filters matter far less for this specific goal.

• A. Very high resolution output only
• B. Fast generation with no account required
• C. Character consistency or reference-image support ✓
• D. A large library of preset art styles

Why: Maintaining a consistent character across multiple images requires features like reference-image conditioning or consistency controls. Resolution, speed, and style presets do not ensure the same character appears across generations.

• A. Cloud tools always produce lower quality images
• B. Local models cannot generate any images at all
• C. Cloud tools offer ease of use but may limit customization and raise data privacy considerations ✓
• D. Open-source models never require a graphics card

Why: Cloud services are convenient and require no setup, but they often restrict customization and involve sending data to a third party. Local open-source models offer more control and privacy at the cost of setup and hardware requirements.

• A. Licensing terms only affect the free tier of a tool
• B. Terms determine whether you own the outputs and whether commercial use is permitted ✓
• C. Licensing only matters for images containing people
• D. Commercial use is always allowed by default with any AI tool

Why: Each generation platform sets its own rules on ownership of outputs and permitted commercial use, and these rules vary widely. Assuming commercial rights without checking can lead to legal problems.

• A. Choosing the tool with the most marketing followers
• B. Relying only on the price listed on the homepage
• C. Testing the tool with a few of their own real prompts and comparing results ✓
• D. Selecting whichever tool was used most recently in a tutorial

Why: Testing with your own representative prompts reveals how a tool actually performs for your use case. Popularity, price alone, or tutorial habits do not reliably predict fit for your specific needs.

• A. Seed
• B. Negative prompt ✓
• C. CFG scale
• D. Sampler

Why: A negative prompt tells the model what to exclude from the image, such as blurriness, extra fingers, or watermarks, helping steer the output away from undesired features.

• A. Makes the image render faster
• B. Increases the image resolution automatically
• C. Makes the output adhere more strictly to the text prompt ✓
• D. Randomizes the composition on each run

Why: CFG scale controls how strongly the model follows the prompt. Higher values increase prompt adherence but can reduce naturalness or introduce oversaturation if pushed too far.

• A. To guarantee the image is in color
• B. To make the output reproducible so the same settings yield the same image ✓
• C. To increase the number of diffusion steps
• D. To remove the need for a prompt

Why: The seed initializes the random noise the model starts from. Keeping it fixed while adjusting other parameters lets you compare changes reliably and reproduce a result you like.

• A. Upscaling
• B. Inpainting only
• C. Control networks such as ControlNet ✓
• D. Tokenization

Why: ControlNet and similar conditioning methods use reference inputs like edge maps, depth maps, or pose skeletons to constrain the spatial structure of the generated image.

• A. It changes the file format of the output
• B. It sets the image's aspect ratio
• C. It guides the visual aesthetic such as brushwork, palette, and mood ✓
• D. It determines how many images are generated

Why: Style descriptors like art movements, mediums, or artist references influence the visual character of the output, including color palette, texture, and rendering technique.

• A. The resolution of the output image
• B. How much the output can deviate from the input image ✓
• C. The number of people detected in the image
• D. The color depth of the final file

Why: Denoising strength determines how much noise is added to the input before regeneration. Higher values allow the model to change the image more substantially, while lower values keep the output closer to the original.

• A. Converting a photo into a painting style
• B. Regenerating only a masked region of an image while keeping the rest unchanged ✓
• C. Increasing the overall brightness of an image
• D. Combining two separate images side by side

Why: Inpainting uses a mask to specify which area should be regenerated. The model fills in the masked region based on the prompt and surrounding context, leaving unmasked pixels intact.

• A. White areas are always ignored by the model
• B. White marks the region to be regenerated, while black marks areas to preserve ✓
• C. White increases the sharpness of the output
• D. White is required for the file format to be valid

Why: The common convention in inpainting is white for the area to edit and black for the area to keep. Understanding this convention is essential to avoid accidentally regenerating the wrong part of an image.

• A. Removing the background from an image
• B. Extending an image beyond its original borders with newly generated content ✓
• C. Printing an image at a larger physical size
• D. Applying a filter to only the edges of an image

Why: Outpainting generates new content outside the original canvas, expanding the scene in one or more directions while trying to match the existing style, lighting, and content.

• A. Using a mask that is much smaller than the object being changed
• B. Making the mask edges slightly feathered and including some surrounding context in the prompt ✓
• C. Setting the image to grayscale before editing
• D. Editing at the lowest possible resolution

Why: Feathered mask edges blend the generated region into the original, and a prompt that describes the surrounding scene helps the new content match lighting, perspective, and style, reducing visible seams.

• A. Whether the image looks visually appealing
• B. Whether the generator's terms of service permit commercial use of the outputs ✓
• C. Whether the image was generated at the highest possible resolution
• D. Whether the prompt used was written in English

Why: Image generators have different terms of service regarding commercial rights. Checking whether outputs may be used commercially is essential to avoid licensing violations, regardless of how good the image looks or what resolution or language was used.

• A. The model was trained without any computing hardware
• B. The training images were collected from the web without permission from their creators ✓
• C. The model was trained only on public domain images
• D. The model was trained by a non-human researcher

Why: Many AI image models were trained on large datasets scraped from the internet, often including copyrighted works whose creators never gave permission. This is a central ethical and legal debate in AI image generation.

• A. Presenting AI images as real photographs to make them more engaging
• B. Clearly labeling images as AI-generated so viewers are not misled ✓
• C. Removing metadata so the image origin cannot be traced
• D. Using AI images only when the subject is a famous person

Why: Transparency about AI generation prevents audiences from being deceived about what is real. Hiding origin or presenting synthetic images as authentic undermines trust and can spread misinformation.

• A. It always violates copyright law in every country
• B. It can enable defamation, harassment, or non-consensual deepfakes that damage the person's reputation or safety ✓
• C. It uses too much computing power
• D. Realistic images of people cannot be produced by current tools

Why: Synthetic images of real people can be used to fabricate false events, harass individuals, or create sexual deepfakes. Using someone's likeness without consent raises serious privacy, dignity, and legal concerns, even if copyright is not the issue.

• A. Publish it immediately since AI outputs are automatically free of rights
• B. Review the tool's license terms, check for similarity to existing works, and seek legal or policy guidance if needed ✓
• C. Assume that because the tool is popular, its outputs are legally cleared
• D. Modify one small detail so the image counts as original work

Why: Responsible users verify the tool's licensing terms, check for resemblance to copyrighted material, and consult appropriate guidance when uncertain. Popularity does not guarantee legal clearance, and minor edits do not automatically resolve rights or ethical issues.