SEAL changes 5% of face pixels to block diffusion identity cloning
On CelebA-HQ and VGGFace2, protected images stay above 41 dB PSNR while identity similarity falls to at most 0.15.
Stories per day · 30 days
7 RSS
Relatedgenerative-models3image-generation3inference3language-models3text-to-image3autoregressive-models2benchmarks2deepfakes2differential-privacy2identity-protection2privacy2time-series2adversarial-perturbation1adversarial-perturbations1agents1anomaly-detection1avatar-refinement1celeba1
76 stories, newest first · one card per story
Stories tagged #diffusion-models
Comfy-Org repackages RelayVSR weights for ComfyUI
The Hugging Face card lists int8 diffusion weights and fp32 Flash VAE files for 2x and 4x upscaling, sourced from kopper/RelayVSR.
Flex-DDCM adapts zero-shot image compression with diffusion guidance
Task guidance is embedded in the compressed trajectory without extra decoding complexity.
Energy-Tweedie links Gibbs noise denoising to Stein scores
The identity extends the Gaussian posterior-mean relation to kernel scoring rules and user-chosen noise paths.
NotaryMark multi-bit watermarking for T2I dataset auditing
NotaryMark aligns watermarks with diffusion latent denoising for robust black-box auditing across datasets and models.
FewStepRL adapts diffusion RL objectives to few-step samplers
The method recasts FlowGRPO, AWM, DiffusionNFT, DiffusionOPSD and ReFL so rollouts no longer need a many-step sampler.
EDGAR Refines Gaussian Avatars via Diffusion
EDGAR corrects artifacts in Gaussian avatar renderings using a fusion block with Sapiens identity features and a Mixture-of-Experts architecture.
StepX-OPD merges distillation stages for faster diffusion models
StepX-OPD lets one student learn all capabilities and few-step sampling using its own rollout states, scoring highest among four-step methods on SD3.5-Medium and FLUX.1-dev while using 37% fewer GPU-hours.
RangeSWA provides training-free length extrapolation for diffusion language models
RangeSWA reduces perplexity and repetition in long-text generation using sliding-window attention and positional remapping.
Paper calibrates diffusion membership inference across two time axes
Reference-Model Calibration uses an earlier checkpoint; Temporal Self-Calibration uses one model's neighboring diffusion timesteps.
SynthFaces dataset enables balanced human generation for fair AI models
SynthFaces expands Arc2Face diffusion model with flow matching to create diverse synthetic faces, reducing privacy risks and biases in real datasets.
Diffusion sampling method improves low-overlap X-ray ptychography
Authors say progressive coupling beats iterative methods and other diffusion samplers when scan overlap is reduced.
Metacognitive controller trims diffusion language-agent inference cost
Across five benchmarks, the plug-in cuts inference FLOPs 20.6% on average versus the strongest efficient baseline while keeping competitive accuracy.
Black-box signatures proposed to detect LoRA adaptations of diffusion models
Tests use only the intrinsic dimension of generated images and assume unknown adaptations and model families.
CLOSURE detects time-series anomalies via consistency and recoverability
A diffusion reconstructor and a relation-closure operator are trained separately and calibrated only on normal data.
LLaCO aligns continuous text diffusion with token competition
An 8B LLaCO model uses spherical noising and semantic retention scheduling and is reported to rival comparable discrete diffusion models.
ConceptACP erases diffusion concepts with sparse autoencoders
On SDXL-Turbo one-step generation, forgetting success rises 36.34 points over the best baseline.
Prior-Consistent DAPS Improves Diffusion Sampling for Nonlinear Inverse Problems
PC-DAPS prunes candidates by prior consistency without ground truth to reduce prior-inconsistent samples on ambiguous nonlinear problems.
Speculative diffusion paper maps joint speed-quality tradeoffs
Authors report that jointly varying diffusion effort and acceptance leniency beats single-step diffusion drafting on the speed-quality frontier.
Orbitol generates crystals constrained to requested space groups
On MP-20, 13.8% of attempts stayed stable or metastable, unique, novel, and in the requested group after relaxation.
Inception claims Mercury 3 beats GPT-6 Luna and Claude Haiku 5.5
Inception's own tests give Mercury 3 eight of 12 wins and 1,068 tokens per second, with no independent verification yet.
3
Block diffusion trades speed for accuracy on longitudinal EHR data
On MIMIC-IV, a block size of 8 evaluates about 1.43 times faster, while linear probes nearly match autoregressive models.
Paper derives exact privacy-utility bounds for diffusion denoisers
Exact two-record sensitivity and an SDP replace loose diameter bounds and isotropic Gaussian noise.
CoRAL pairs a diffusion planner with an autoregressive executor
A diffusion planner drafts K parallel reasoning paths; a large autoregressive model selects one and writes the answer.