Objective, representation, and sampler
A generative system has three separable choices. The representation defines what the model sees: raw pixels, compressed latents, audio codes, text tokens, or trajectories. The training objective defines what it learns to predict. The sampler turns that learned distribution into an output. A faster sampler can improve usability without changing training; a better representation can reduce cost while discarding detail.
Mental model: generation is controlled sampling. Judge the learned distribution, the control interface, and the sampling cost separately.
Autoregressive and iterative generation
Autoregressive models predict the next element conditioned on previous elements. They naturally support variable-length sequences but generate sequentially. Diffusion-style models learn to reverse a corruption process and typically refine a whole representation over multiple steps. Flow and consistency approaches seek more direct transformations. Architecture labels matter less than measured quality, control, and latency on the target medium.
Conditioning is where user intent enters
Text prompts, reference images, masks, layouts, audio, camera paths, and control signals constrain output. Strong prompt adherence can conflict with diversity or realism. Guidance strength often moves along that tradeoff. Evaluation should use compositional prompts, rare combinations, spatial relationships, and negative constraints—not only broad aesthetic requests.
One sample proves almost nothing
Cherry-picked outputs conceal variance. Compare multiple random seeds, fixed prompts, and failure cases. Human preference studies should randomize presentation, hide model identity, define criteria, and report agreement. Automated similarity or quality metrics capture limited aspects and can be gamed by the same biases present in their feature models.
Memorization and provenance matter
Large training corpora can contain duplicates, copyrighted material, personal data, or benchmark examples. Similarity checks are imperfect but should examine near-duplicates and recognizable training items. Output provenance mechanisms can help identify generated media, yet they must survive common transformations and should not be treated as a complete defense against misuse.
World models need predictive usefulness
A generated video that looks plausible is not necessarily a useful simulator. For planning or robotics, the model must preserve action-conditioned consequences, object permanence, geometry, and uncertainty across time. Evaluate whether decisions made inside the model transfer to the real environment, not only whether samples look coherent.
Questions to ask when reading a paper
- What changed? Data, representation, objective, architecture, conditioning, or sampler?
- How were examples selected? Fixed prompt set, random seeds, and visible failure cases?
- What is the speed-quality point? Steps, resolution, duration, hardware, memory, and latency?
- How is control measured? Composition, identity, temporal consistency, and instruction following?
- What data risks remain? Memorization, consent, provenance, bias, and unsafe capabilities?
A useful comparison protocol
Choose a fixed prompt suite covering simple, compositional, and adversarial requests. Generate several seeds per prompt with declared settings. Record runtime and memory. Have reviewers score adherence, technical defects, and preference as separate dimensions. Publish representative failures alongside successes. For editing or control, measure how much unrelated content changes when one condition changes.
Bottom line
Generative research is strongest when it replaces the highlight reel with a distribution: repeated samples, explicit controls, cost, failure modes, data provenance, and evaluation matched to the intended use.