Get Started
Topic · 18 recaps
Diffusion Models
Generative models that learn to reverse a noising process. The dominant approach behind modern image, video, and audio synthesis.
Play all
...
Posts
Questions
Home
Topics
Search
Library
Questions researchers are working on
Follow a question through Rcap’s explanations and the latest papers addressing it.
Search
Can causal fairness constraints transfer across synthetic-data generators and privacy levels without sacrificing fidelity?
Synthetic data releases must suppress unfair causal pathways while retaining enough statistical structure for downstream use. It is unclear whether these controls remain effective when the generator family or formal privacy guarantee changes.
Do gains from fine-tuned genomic variant-effect prediction transfer to zero-shot functional sequence inpainting?
Genomic models can score variants after supervised fine-tuning yet fail to reconstruct biologically functional sequence in zero-shot settings. This leaves unclear whether a masking prior that helps one objective captures the constraints needed for the other.
How can attackers evade black-box AIGC detectors using a frozen diffusion model without source images or detector-aware retraining?
AIGC detectors must identify synthetic images even when generation can be adjusted to reduce detectable signals. This is difficult when the attacker can observe only detector responses and cannot alter the generator or rely on source images.
How can attention heads be pruned in text-to-image diffusion transformers without losing prompt-specific object identity?
During denoising, semantic information may be maintained by structural template tokens and image-to-text interactions rather than by the prompt tokens that initially encode it. This makes it difficult to identify redundant attention computation without disrupting object identity.
How can audio-video diffusion models preserve intended conditioning when biased cross-modal attention reroutes semantics?
In audio-video diffusion generation, cross-attention among text, audio, and video can route semantics bidirectionally rather than respecting intended conditioning. Learned biases may cause one modality to override prompts, producing visually canonical but semantically incorrect outputs.
How can autonomous-driving planners be stress-tested in realistic closed-loop scenarios that expose failures missed by nominal benchmarks?
Nominal benchmark performance can miss failures that emerge when traffic participants create rare, interacting hazards. Closed-loop evaluation needs scenarios that remain realistic while probing those behaviors.
How can autonomous-driving scenario generators reliably induce collisions at a requested region of the target vehicle?
Existing autonomous-driving scenario generators can produce crashes but offer limited control over where the target vehicle is struck. This makes it difficult to construct tests for region-specific collision behavior.
How can autoregressive LLM decoding generate multiple tokens in parallel at large batch sizes without sacrificing quality?
Autoregressive LLMs repeatedly predict one next token at a time, limiting throughput and increasing serving costs as batch sizes grow. Parallelizing generation is difficult because it must retain the quality and distribution of the original autoregressive model.
How can causal streaming video editing remain real-time while preserving backgrounds and unedited regions over long sequences?
Each frame must be edited without access to future frames, while backgrounds and untouched regions remain stable over time. Meeting that consistency requirement without sacrificing interactive responsiveness is difficult.
How can cloud-cover forecasts extend beyond nowcasting despite imperfect NWP initialization while preserving satellite spatial detail?
Satellite observations provide a useful initial cloud state, but cloud fields form, dissipate, and deform, causing short-range forecast skill to decline after the first few hours. Longer horizons must represent atmospheric evolution without losing the observed spatial structure when NWP initialization differs from the satellite state.
How can conditioned diffusion models be guided without score functions when diffusion coefficients are singular and conditional densities are absent or nonsmooth?
Score-based guidance becomes difficult when the diffusion coefficient is singular and relevant conditional densities are unavailable or lack smoothness. Conventional density-derived guidance may therefore not be well defined in these processes.
How can continuous diffusion language models reduce denoising steps without sacrificing text-generation quality?
Continuous diffusion language models generate text through repeated denoising steps, which can make inference too slow for practical use. Reducing the number of steps may also degrade the quality of generated language.
How can continuous image-generation controls achieve measurable scales and avoid interference when composed?
Slider changes may not produce consistent, observable changes in the rendered image, and separate controls can interfere when combined. This makes it difficult to determine which attributes are controllable, compare control strengths, and predict composed results.
How can cross-domain image compositing preserve foreground identity while matching the background’s visual style?
Different-domain images cannot be reconciled reliably through tone adjustment alone. Stylizing the inserted object to fit its surroundings can also alter the object’s recognizable identity.
How can DDPM sampling error be controlled for variance schedules beyond reverse-SDE discretizations?
DDPM sampling is commonly analyzed by discretizing its reverse SDE, which can restrict the variance schedules and parameter choices considered. Relating DDPM sampling to the discretization of a Föllmer process may provide a broader framework for understanding these choices and their resulting sampling errors.
How can diffusion image and video generators be preference-aligned without inefficient training exploration or inference-time search?
Policy-gradient preference optimization can explore inefficiently, become trapped in local optima, and exploit reward signals in ways that reduce semantic faithfulness or visual realism. Alignment must also avoid requiring extra search during generation.
How can diffusion language models adapt infilling span lengths without sacrificing quality or inference speed?
Infilling quality depends on the missing span's length, yet diffusion language models typically require that length before generation. Existing adaptive approaches can be sensitive to their starting length or add costly length changes and repeated searches.
How can diffusion language models support reliable mobile-edge agents under tight latency and resource constraints?
Mobile-edge agents must generate and act despite limited compute, memory, energy, and network capacity while meeting privacy and reliability requirements. Diffusion language models refine multiple tokens with bidirectional context, creating latency and quality trade-offs that differ from sequential decoding.
How can diffusion models forecast multivariate time series reliably when variables differ in noise and observations are corrupted?
Multivariate series can contain variables with different evolutionary patterns and noise levels, making shared conditioning signals prone to trusting unreliable observations. Corruption can further distort the resulting probabilistic forecasts.
How can diffusion models preserve fine perceptual detail when reconstructing high-resolution images from low-resolution inputs?
Pixel-domain noise-prediction objectives can produce oversmoothed super-resolved images, making fine structure difficult to recover from low-resolution observations.
Previous
1 / 6
Next