Generative AI Engineer interview questions

They parameterize the same denoising problem differently, but redistribute target scale and error across noise levels.

• Epsilon prediction estimates added noise and is simple, yet its effective emphasis can be uneven at very high and low signal-to-noise ratios.
• x0 prediction estimates the clean sample directly, which can aid reconstruction but becomes numerically demanding when the input is mostly noise.
• v prediction combines clean-sample and noise terms so target variance is better balanced across common schedules, but the scheduler and inference conversion must match the checkpoint.

SNR-based weighting controls how much each noise level contributes instead of letting easy or high-magnitude timesteps dominate the update.

• The signal-to-noise ratio identifies whether a timestep is mainly reconstruction-like or noise-prediction-like under the chosen schedule.
• Min-SNR weighting clips or caps the influence derived from SNR, reducing conflict between gradients from very different timesteps.
• The formula depends on epsilon, x0, or v prediction, so I would validate the implementation and compare convergence plus prompt-aligned quality, not copy one gamma blindly.

DDPM sampling follows a stochastic reverse Markov chain, while DDIM can use a deterministic non-Markovian path for the same trained model.

• DDPM injects variance at reverse steps, so repeated runs can diverge even from related intermediate states.
• DDIM exposes an eta-style control where zero gives a deterministic path for a fixed initial latent and larger values restore stochasticity.
• DDIM often reaches usable quality in fewer steps, but speed and quality still depend on the checkpoint, timestep spacing, and implementation.

I would compare them on the exact model, schedule, step budget, guidance range, and target prompt distribution because sampler rankings do not transfer reliably.

• Euler methods are simple and can produce strong results at modest step counts, while ancestral variants deliberately add noise and increase run-to-run diversity.
• DPM-Solver and DPM++ families use higher-order or data-prediction formulations that can be efficient, but multistep history and schedule choice affect stability at low step counts.
• I would hold prompts, seeds, and the number of outputs per prompt constant, then measure latency, failure rate, preference wins, and diversity at several step budgets.

They reduce artifacts caused by applying one large classifier-free guidance scale uniformly throughout denoising.

• High CFG can improve prompt adherence but overexpose colors, collapse variation, or push the sample away from the model's trained distribution.
• CFG rescale adjusts the guided prediction magnitude toward the conditional prediction's statistics, limiting excessive contrast without simply removing guidance.
• Dynamic schedules vary guidance by timestep, often weakening it where conditioning is unhelpful, and must be tuned with both adherence and visual-quality slices.

Flow matching learns a time-dependent velocity field that transports samples between noise and data distributions, rather than directly fitting a denoising score objective.

• The training target comes from a chosen probability path and its conditional velocity, with straight interpolation being a common simple path.
• Rectified flow is flow matching on a straight interpolation between paired source and data samples; the learned marginal ODE is not guaranteed to be straight, while reflow can progressively straighten its coupling and reduce integration error.
• The path, parameterization, data-noise coupling, and ODE solver still determine performance, so the label alone does not guarantee one-step generation.

A latent autoencoder moves denoising into a smaller spatial and channel representation, cutting attention and convolution cost compared with pixel-space diffusion.

• The encoder compresses an image to latents and the decoder reconstructs pixels after sampling, so its downsampling factor largely sets the compute saving.
• Aggressive compression can erase small text, faces, textures, or high-frequency detail before the diffusion model sees them.
• I would evaluate reconstruction separately from generation because decoder artifacts and latent scaling errors cannot be fixed by changing the sampler.

DiT applies transformer blocks to latent image patches, making generative capacity easier to scale with depth, width, tokens, and compute.

• Patchifying latents turns spatial generation into a token sequence while timestep and class or text conditioning modulate transformer blocks.
• Transformer scaling benefits from mature attention kernels and parallel training infrastructure, but token count still grows quickly with resolution.
• More parameters or FLOPs are not sufficient by themselves, so I would track quality against training compute, data scale, autoencoder quality, and inference cost.

MM-DiT lets text and image latent tokens participate in joint attention while retaining modality-specific representations or projections.

• Text and image streams can use separate weights for parts of the block, then exchange information through a shared attention operation.
• Unlike one-way cross-attention, joint attention allows image tokens to influence text-token states as well as receive text conditioning.
• Sequence length, masking, positional encoding, and modality balance become design constraints because both token sets consume the same attention computation.

LCM distillation trains a student to map points on a diffusion trajectory toward a consistent clean endpoint, so inference can skip most teacher steps.

• Training uses teacher-derived transitions or a probability-flow trajectory and enforces consistency between appropriately paired noisy states.
• A boundary condition anchors the consistency function near the data endpoint, preventing a trivial arbitrary mapping.
• One to eight steps can be practical, but guidance distillation, scheduler compatibility, and prompt-distribution coverage determine how much quality survives.

Progressive distillation repeatedly halves the required denoising steps, while adversarial distillation can train a very low-step student using a discriminator or learned realism signal.

• Progressive stages provide a controlled teacher-student path, but errors can accumulate after several rounds and every stage adds training cost.
• Adversarial objectives sharpen perceptual realism at one or a few steps, yet may reduce diversity or introduce GAN-like instability and artifacts.
• I would compare prompt fidelity, diversity, rare-concept failures, and latency at the exact deployment step count rather than report only FID.

ControlNet learns a parallel conditioning branch from a copied backbone and injects its features into a mostly frozen denoiser.

• A pose, edge, depth, or segmentation map enters the control branch at matching spatial resolutions.
• Zero-initialized connections make the added residuals start near zero, preserving the pretrained model at the beginning of training.
• Control strength and start or end timesteps must be tuned because excessive residuals can overpower text conditioning or reproduce detector noise.

IP-Adapter encodes a reference image into separate conditioning tokens and injects them through added attention paths alongside text conditioning.

• A vision encoder and projection network convert the reference into tokens compatible with the denoiser's attention dimensions.
• Separate image cross-attention weights let the base text pathway remain frozen and keep the adapter relatively small.
• Adapter scale balances identity or style transfer against prompt adherence, and face-specific or composition-specific variants require matching evaluation data.

The adapters can compete for the same attention or residual capacity, so individually good settings may fail when their effects are summed.

• A strong ControlNet can lock geometry while an IP-Adapter tries to impose reference composition, producing distorted subjects or ignored text.
• Scales, timestep windows, injection sites, and normalization differ across adapters, so composition needs a joint grid rather than isolated tuning.
• I would test every supported combination and order on conflict prompts, then cap incompatible pairs or route them to a specialized pipeline.

Rank sets the update capacity, alpha scales the low-rank update, and target modules decide where that capacity can change the model.

• A higher rank can represent richer changes but increases trainable parameters, optimizer memory, file size, and overfitting risk.
• Many implementations apply an effective scale related to alpha divided by rank, so changing rank without checking scaling changes optimization behavior.
• Cross-attention projections are common targets, while adding self-attention, feed-forward, convolutional, or text-encoder modules should follow a measured capability gap.

I would merge a validated fixed combination for simpler and faster serving, but keep runtime composition when users need independent adapter selection or scaling.

• Merging removes adapter switching overhead and can simplify export, yet the resulting weights are tied to one base checkpoint and one set of merge coefficients.
• Runtime composition preserves reversibility and per-request scales, but extra adapter reads or kernels consume memory bandwidth and complicate batching.
• Multiple updates are not guaranteed to add cleanly, so both paths need conflict prompts and regression checks before deployment.

Prior preservation mixes instance examples with class examples so learning a subject does not collapse the broader class around that subject.

• Instance loss teaches a rare identifier to reproduce the target person or object from a small dataset.
• Class images and a class prompt provide a prior loss that discourages language drift and loss of diversity for ordinary class prompts.
• The loss weight, class-image quality, duplicates, and subject coverage need tuning because an excessive prior can weaken identity learning.

Text-encoder tuning can bind a new concept more strongly, but it can also distort the language space used by every other prompt.

• With few captions, token embeddings may absorb background, pose, or style correlations and cause prompt leakage.
• Updating a large text encoder adds optimizer memory and can overfit faster than the denoiser, especially with an aggressive learning rate.
• I would start frozen, then test a lower-rate partial update only if identifier binding remains weak, with general-prompt and compositional regressions.

I would curate for correct visual-text alignment, coverage of intended variation, and removal of shortcuts the model could memorize.

• Deduplicate exact and near-duplicate assets, remove corrupt or low-resolution files, and balance subjects, styles, poses, backgrounds, and hard cases.
• Captions should name attributes the model must control while avoiding confident labels for invisible details or boilerplate copied across every image.
• I would version source rights, filtering rules, captioner and human edits, then inspect stratified samples rather than trust only automated scores.

Aspect-ratio bucketing groups similarly shaped images into batches so resizing wastes less content and padding wastes less compute.

• Each bucket uses dimensions compatible with the autoencoder and patch stride while keeping pixel or token budgets bounded.
• Sampling only by raw dataset frequency can overrepresent common shapes, so bucket probabilities may need balancing.
• Too many buckets fragment batches and hurt throughput, while too few force crops or distortions, so I would profile utilization and composition quality together.