Generative Model Papers

Generative model paper note는 sample, transform, denoise, decode, structured object generation을 다룹니다.

이 선반의 기준은 “새 sample을 만들거나 distribution을 모델링하는가”입니다.

x \sim p_{θ} (x ∣ c)

where $c$ can be text, class label, protein pocket, scaffold, structure constraint, trajectory time, or another conditioning signal.

Reading Axes

어떤 distribution을 모델링하는가?
generation 방식이 autoregressive, diffusion-based, flow-based, energy-based, latent-variable 중 무엇인가?
어떤 objective를 optimize하는가?
validity, diversity, novelty, task utility를 어떻게 평가하는가?
model이 text, molecule, protein, structure, image, action 중 무엇을 생성하는가?
claim이 likelihood, sample quality, controllability, efficiency, downstream utility 중 무엇에 관한 것인가?

Strongest Claim	Put Here?	Cross-Link
new diffusion, flow, autoregressive, latent, or energy objective	yes	Generative models
molecule or protein generator	yes, if generation is the core contribution	Computational Biology papers
docking or pose search with generative component	sometimes	Structure-based modeling papers
architecture block for many tasks	no, unless generation evidence is central	Architecture papers
LLM decoding or instruction generation	usually no	Architecture papers

Generation papers often report many metrics that answer different questions.

Metric Family	Asks	Common Trap
validity	is the output syntactically/physically valid?	invalid denominator removed after filtering
diversity	are samples different from each other?	high diversity with low utility
novelty	are samples unlike training data?	novelty defined by weak similarity threshold
quality	does a generated sample look or score well?	evaluator model shares training bias
controllability	does condition $c$ change output as intended?	condition leakage or easy constraints
downstream utility	does generated output improve a task?	cherry-picked candidates or missing baseline
efficiency	is sampling cheaper or faster?	hardware, batch size, and sampler budget hidden

For each paper, separate:

objective \to sampler \to filter \to metric \to claim

Field	Question
Objective	likelihood, denoising, score, velocity, energy, reward, reconstruction?
Sampler	how many steps, candidates, retries, guidance, temperature, or constraints?
Filter	are invalid or low-quality samples removed before scoring?
Condition	what information is available at generation time?
Baseline	same budget, same filter, same evaluator, same dataset?
Utility	is the generated object useful under an independent check?