Diffusion Model

A Diffusion Model is a class of probabilistic generative model that learns to synthesise data by reversing a learned forward diffusion process in which training examples are progressively corrupted with Gaussian noise across a fixed Markov chain of timesteps. At inference time the model iteratively denoises a sample drawn from pure noise, guided by a parametrised score function or noise-prediction network, until a high-fidelity output is recovered. Architecturally, the denoising backbone is typically a U-Net or Vision Transformer conditioned on timestep embeddings and optional guidance signals such as text or class labels. Diffusion models achieve state-of-the-art quality on image, audio, video, and molecular generation tasks and underpin production systems including Stable Diffusion, DALL-E 3, Sora, and AudioLDM.

Semantic Classification

Content

Compositional Relationships (Components)

SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:hasPart ai:ForwardDiffusionProcess))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:hasPart ai:ReverseDiffusionProcess))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:hasPart ai:NoiseSchedule))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:hasPart ai:ScoreFunction))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:hasPart ai:DiffusionTransformer))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:hasPart ai:ClassifierFreeGuidance))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:hasPart ai:LatentEncoder))
SubClassOf(ai:LatentDiffusionModel
  ObjectSomeValuesFrom(ai:hasPart ai:VariationalAutoencoder))

Dependency Relationships

SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:requires ai:TrainingData))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:requires ai:GPUCompute))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:requires ai:LatentSpace))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:requires ai:Backpropagation))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:dependsOn ai:MarkovChain))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:dependsOn ai:StochasticDifferentialEquation))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:dependsOn ai:GaussianDistribution))

Capability Relationships

SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:enables ai:ImageGeneration))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:enables ai:TextToImage))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:enables ai:VideoGeneration))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:enables ai:AudioSynthesis))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:enables ai:DrugDiscovery))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:enables ai:ProteinStructureDesign))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:enables ai:SyntheticDataGeneration))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:enables ai:MedicalImagingSynthesis))

Implementation Relationships

SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:implements ai:ScoreMatching))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:implements ai:VariationalLowerBound))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:implements ai:DenoisingObjective))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:uses ai:UNet))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:uses ai:AttentionMechanism))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:uses ai:ClassifierFreeGuidance))

Reduction Relationships

SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:reducesTo ai:GenerativeModel))
SubClassOf(ai:LatentDiffusionModel
  ObjectSomeValuesFrom(ai:reducesTo ai:DiffusionModel))
SubClassOf(ai:RectifiedFlowModel
  ObjectSomeValuesFrom(ai:reducesTo ai:DiffusionModel))
SubClassOf(ai:ScoreBasedGenerativeModel
  ObjectSomeValuesFrom(ai:reducesTo ai:DiffusionModel))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:contrastsWith ai:GenerativeAdversarialNetwork))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:contrastsWith ai:VariationalAutoencoder))
SubClassOf(ai:DiffusionModel
  ObjectSomeValuesFrom(ai:relatedTo ai:StochasticDifferentialEquation))
SubClassOf(ai:DiffusionTransformer
  ObjectSomeValuesFrom(ai:reducesTo ai:DiffusionModel))

About

Diffusion models constitute a probabilistic generative paradigm that has risen from relative obscurity to dominate practical generative AI across modalities in under five years. The core mathematical intuition is to define a forward process that gradually destroys the structure of real data by adding Gaussian noise, and then to train a neural network to reverse this destruction step by step. Because the forward process is analytically tractable — the marginal q(x_t | x_0) at any arbitrary timestep t can be computed in closed form without iterating through all intermediate steps — training is efficient via a simple denoising mean-squared-error loss on the predicted noise. The reverse process, approximated by the trained network, reconstructs data from noise by iterating the learned denoising steps from t = T (pure noise) down to t = 0 (clean sample).

The framework was formalised as Denoising Diffusion Probabilistic Models (DDPMs) by Ho, Jain, and Abbeel (NeurIPS 2020), building on the score-matching literature of Hyvärinen (2005) and the Langevin dynamics samplers studied by Song and Ermon (NeurIPS 2019). The decisive practical breakthrough came with Rombach et al.’s Latent Diffusion Models (CVPR 2022), which moved the diffusion chain from pixel space into the compressed latent space of a pretrained Variational Autoencoder, reducing the spatial dimensionality by 4–16× and enabling high-resolution synthesis on consumer hardware. This architecture forms the basis of Stable Diffusion and its successors. A parallel line of theoretical work by Song, Sohl-Dickstein, Kingma, Kumar, Ermon, and Poole (ICLR 2021) reformulated diffusion models in the continuous-time language of Stochastic Differential Equations, unifying DDPM (VP-SDE), SMLD (VE-SDE), and sub-VPSDE under a single framework and enabling more flexible training and sampling.

The period 2023–2026 has been defined by the architectural shift from U-Net backbones to Diffusion Transformer (DiT) backbones combined with Rectified Flow or Flow Matching training objectives. The DiT architecture, introduced by Peebles and Xie (ICCV 2023), treats the image as a sequence of patch tokens processed by a transformer with adaptive layer normalisation conditioned on timestep and class embeddings. Stable Diffusion 3 (Esser et al. 2024) introduced the Multimodal Diffusion Transformer (MMDiT) which applies separate sets of weights for text and image tokens and allows bidirectional attention between modalities. The FLUX family (Black Forest Labs, 2024–2025) — FLUX.1, FLUX 1.1 Pro, and FLUX 2 (32B parameters, November 2025) — implements this architecture at scale, achieving near-photorealistic quality with 4.5-second generation times on professional hardware. By late 2025, MLCommons introduced FLUX.1 as the benchmark workload in MLPerf Training Round 5.0, cementing its position as the standard reference architecture.

Formal Algorithm

Forward Process (q):

• Define q(x_t | x_{t-1}) = N(x_t; √(1-β_t) x_{t-1}, β_t I) for a noise schedule β₁, …, β_T.

• The marginal is q(x_t | x_0) = N(x_t; √ᾱ_t x_0, (1-ᾱ_t) I) where ᾱ_t = Π_{s=1}^t (1-β_s).

• This closed-form marginal allows direct sampling of x_t from x_0 without iterating intermediate steps.

  Reverse Process (p_θ):

• Model p_θ(x_{t-1} | x_t) = N(x_{t-1}; μ_θ(x_t, t), Σ_θ(x_t, t)).

• Training objective: L = E_{x_0,ε,t}[‖ε - ε_θ(√ᾱ_t x_0 + √(1-ᾱ_t) ε, t)‖²], minimising MSE between predicted and actual noise.

• At inference: iteratively apply p_θ(x_{t-1} | x_t) from t=T down to t=0.

  Score Formulation (Song et al. 2021):

• The denoising network implicitly learns the score: ε_θ(x_t, t) ≈ -√(1-ᾱ_t) ∇_{x_t} log q(x_t).

• Langevin MCMC sampling follows: x_{t-1} = x_t + η ∇ log p(x_t) + √(2η) z, where z ~ N(0,I).

  Rectified Flow (Liu et al. 2022; Esser et al. 2024):

• Defines straight-line trajectories between noise and data: x_t = (1-t)x_0 + t ε.

• Training objective: L = E_{t,x_0,ε}[‖(x_0 - ε) - v_θ(x_t, t)‖²], learning the velocity field v_θ.

• Near-linear trajectories enable accurate ODE integration with very few function evaluations.

  Components / Architecture

  A complete modern latent diffusion system comprises five interacting components:

• Encoder-Decoder (VAE or Tokeniser): Compresses inputs from pixel/token space into a compact latent representation. In SDXL the VAE operates at 8× spatial downsampling; in SD3/FLUX, continuous latent tokens enable higher compression ratios. The decoder inverts the latent at inference time.

• Denoising Backbone (U-Net or DiT): The primary parametric component. U-Net variants use hierarchical encoder-decoder with skip connections and cross-attention for conditioning. Diffusion Transformer (DiT/MMDiT) variants process image patches as sequences with full self-attention, enabling better scalability and text understanding. FLUX 2’s 32B parameter MMDiT architecture applies separate learned projections for image and text tokens with parallel cross-modal attention.

• Noise Schedule: Determines the variance trajectory β_t (or, in flow matching, the interpolation path). Linear schedules (DDPM) have been largely superseded by cosine (Improved DDPM), log-linear (SDXL), and logit-normal (SD3) schedules that provide more uniform loss weighting across timesteps.

• Classifier-Free Guidance (CFG): During training, the conditioning signal c (e.g., text embedding from CLIP or T5) is randomly dropped with probability p (typically 10–20%), training both conditional and unconditional branches jointly. At inference: ε̂ = ε_θ(x_t, ∅) + w(ε_θ(x_t, c) - ε_θ(x_t, ∅)) where w is the guidance scale (typical values 3–12).

• Conditioning Encoder: Projects the conditioning signal into the backbone. CLIP ViT-L/14 encodes text for cross-attention in SDXL; T5-XXL and CLIP are used jointly in SD3/FLUX for richer semantic understanding. IP-Adapter provides image-prompt conditioning via decoupled cross-attention. ControlNet adds zero-convolution spatial conditioning branches (depth, edge, pose) that bypass the frozen backbone.

• Sampler (ODE/SDE Solver): Implements the discrete-time reverse process. DDPM uses ancestral sampling (T=1000 steps). DDIM (Song et al. 2020) enables deterministic non-Markovian sampling at 50–100 steps. DPM-Solver (Lu et al. 2022), PNDM, and UniPC reduce steps to 10–25. Latent Consistency Models (LCM, Song et al. 2023) and Turbo Diffusion distillation achieve 1–4 steps by learning consistency functions. Flow matching / rectified flow enables high-quality generation with 20 ODE steps.

  Use Cases / Major Families

  Image Generation and Editing: Stable Diffusion (all variants), Midjourney, and FLUX dominate professional and consumer workflows in 2026. FLUX 2 (32B, Black Forest Labs) leads on photorealism and aesthetic quality. GPT Image 1.5 (OpenAI, launched April 2026) replaced DALL-E 3 and claims the top position on Image Arena leaderboards with a 93% win rate. Ideogram 3.0 (March 2025) excels at text rendering with ~90–95% typography accuracy. Inpainting (SDEdit), semantic editing (Prompt-to-Prompt), and subject-preservation (IP-Adapter, DreamBooth) extend base models to controllable editing workflows.

  Video Generation: Sora (OpenAI, 2024–2025) pioneered text-to-video at 1080p with reasonable physics simulation before OpenAI deprecated the standalone Sora application in March 2026, transitioning capabilities into ChatGPT. Sora 2.0 features “Physics Compliance Mode” using physics simulation pre-rendering. Runway Gen-4 (2025), Kling (Kuaishou), and Pika Labs provide competing video generation offerings, and 2026 sees widespread commercial use in B-roll, advertising, and music videos. The technical basis is typically a spatiotemporal U-Net or 3D DiT processing video as a sequence of latent frames, trained on large-scale licensed video datasets.

  Audio Synthesis: AudioLDM (2023), Stable Audio (Stability AI), Tango, and AudioLDM 2 apply latent diffusion over mel-spectrogram or continuous audio representations, generating music, speech, and sound effects from text descriptions. MusicGen (Meta) is autoregressive but coexists with diffusion-based audio generators. DiffWave (Kong et al. 2021) is the foundational diffusion model for raw waveform synthesis.

  Drug Discovery and Molecular Generation: RFdiffusion (Watson et al., Science 2023, Baker Lab) generates novel protein backbones by reverse-diffusing over SE(3) space of rotation and translation, enabling design of proteins with specific binding or structural properties not found in nature. DiffSBDD generates small-molecule ligands conditioned on binding pocket geometry. DiffMC-Gen (2025) addresses multi-conditional molecular generation with diverse physicochemical properties. These tools are actively being applied in UK pharmaceutical research at AstraZeneca (Cambridge), GSK, and academic drug discovery centres at UCL, Edinburgh, and Manchester.

  Medical Imaging: Conditional diffusion models synthesise rare pathology cases (e.g., glioblastoma MRI) for Data Augmentation of training datasets where real examples are scarce. CT-to-MRI cross-modal synthesis, MRI acceleration (synthesising fully-sampled images from undersampled k-space), and CT denoising are active clinical applications. The NHS AI Lab and NHSX have explored diffusion-based data augmentation to address long-tail class imbalance in diagnostic AI.

  3D and Spatial Computing: Point-E and Shap-E (OpenAI) apply diffusion to point clouds and implicit NeRF/3DGS parameters. Zero123 and SyncDreamer generate 3D views from single images. Generative models for 3D assets bridge into Spatial Computing and Digital Twin workflows, enabling AI-generated virtual environments for XR applications.

  Synthetic Data Generation: Diffusion-generated images, video, and sensor data augment training pipelines for robotics, autonomous vehicles, and scientific simulation. Differential Privacy-constrained diffusion models enable privacy-preserving Synthetic Data generation, producing statistically realistic data without exposing individual records.

  Academic Context

  The intellectual ancestry of diffusion models spans thermodynamics (non-equilibrium statistical mechanics), probabilistic graphical models, and variational inference. The connection between score functions and generative modelling was established by Hyvärinen (2005, “Estimation of Non-Normalized Statistical Models by Score Matching”) and extended to denoising score matching by Vincent (2011). Song and Ermon’s NCSN model (NeurIPS 2019, “Generative Modeling by Estimating Gradients of the Data Distribution”) demonstrated high-quality image generation via annealed Langevin dynamics driven by a learned score network. Ho, Jain, and Abbeel’s DDPM (NeurIPS 2020) connected these approaches to the latent variable model formulation of Sohl-Dickstein et al. (ICML 2015, “Deep Unsupervised Learning using Nonequilibrium Thermodynamics”), providing a training framework that achieves competitive sample quality with a simple MSE loss. Song et al.’s SDE framework (ICLR 2021) provided the unifying theoretical umbrella. The DALL-E 2 (Ramesh et al. 2022) and Imagen (Saharia et al. 2022) papers demonstrated that text-conditioned cascaded diffusion at scale achieves photorealistic generation. The latent diffusion reformulation (Rombach et al. 2022) enabled Stable Diffusion as an open-weights accessible system. The architectural transition to DiT (Peebles & Xie 2023) and MMDiT (Esser et al. 2024) completed the shift from convolutional to transformer backbones. Flow Matching (Lipman et al. 2023; Albergo & Vaitl 2023) and Rectified Flow (Liu et al. 2022) provided cleaner training objectives that have been adopted by SD3 and FLUX.

  Key academic groups include: Berkeley AI Research (BAIR) — Ho, Jain, Abbeel, Song (DDPM, SDE formulation); CMU — Ermon, Song (NCSN, score matching); Heidelberg / Stability AI — Rombach, Blattmann (LDM, SDXL); MIT — Salimans; Oxford, Cambridge, and UCL in the UK.

  Current Landscape (2026)

  The 2025–2026 landscape is characterised by the dominance of DiT + Rectified Flow architectures and the maturation of video generation into professional creative workflows. FLUX 2 (Black Forest Labs, 32B, November 2025) leads professional image generation with multi-reference conditioning supporting up to 10 reference images for identity and style preservation. OpenAI deprecated DALL-E 3 in May 2026 and transitioned to GPT Image 1.5 (gpt-image-2 model), which achieved the top position on Image Arena leaderboards with a record-breaking +242 point lead. Ideogram 3.0 remains the leader for text-in-image synthesis with ~90–95% typography accuracy. Midjourney v7 competes on aesthetic quality and community-oriented generation workflows.

  Video generation has moved from experimental to commercial: Runway Gen-4, Sora 2.0 (within ChatGPT), and Kling (Kuaishou) are used in commercial advertising, music video production, and B-roll for film. OpenAI deprecated the standalone Sora application in March 2026 but continues operating video generation via ChatGPT Pro. The DALL-E API retirement illustrates broader market consolidation around foundation model APIs rather than task-specific endpoints.

  MLCommons introduced FLUX.1 as an MLPerf Training benchmark in October 2025 (Training Round 5.0), indicating that diffusion model training is now considered a standardised performance measurement workload alongside LLM training. Content provenance has become a regulatory priority: C2PA (Coalition for Content Provenance and Authenticity) cryptographic manifests are embedded in images generated by Adobe Firefly, DALL-E 3, and several Stability AI products. Google DeepMind’s SynthID watermarking is integrated into Imagen and Gemini image generation APIs.

  In drug discovery, the impact of diffusion-based protein design has been concrete: RFdiffusion has been used to design proteins with novel binding properties validated in wet lab experiments, with implications for UK pharmaceutical firms including AstraZeneca and GSK. The 2025 review “Diffusion Models at the Drug Discovery Frontier” (Biology, MDPI) synthesises the state of the field including small molecule generation and therapeutic peptide design.

  Inference efficiency remains a key engineering challenge. Latent Consistency Models, Turbo Diffusion, and Flow Matching distillation methods enable high-quality generation in 1–4 steps. Caching mechanisms such as DiCache (2025, arXiv:2508.17356) and LiteAttention (2025, arXiv:2511.11062) accelerate DiT inference by reusing attention computations across timesteps.

  UK Context

  The UK hosts a distinctive combination of academic research, pharmaceutical industry, and creative industry engagement with diffusion models. Imperial College London contributes through its machine learning and computer vision groups, with researchers exploring conditional diffusion for medical imaging and scientific data synthesis. University of Edinburgh has strong expertise in generative modelling and probabilistic inference, with the School of Informatics hosting researchers who have contributed to the theoretical foundations of score-based models and variational methods. University College London (UCL) contributes via the Centre for Artificial Intelligence and healthcare informatics groups; UCL’s connections to NHS trusts and UCLH have enabled clinical validation studies of diffusion-based medical imaging tools.

  University of Manchester and Manchester Metropolitan University host research into AI for creative industries, with diffusion model applications in digital art, fashion design, and cultural heritage preservation. University of Leeds contributes through its Data Analytics Institute, exploring diffusion models for scientific simulation in materials science and climate modelling. Newcastle University and the National Innovation Centre for Data (NICD) are exploring diffusion-based synthetic data generation for privacy-preserving analytics in the north of England.

  Commercially, Stability AI (London-headquartered until 2024 restructuring) was the most prominent UK-origin diffusion model company, having released the open-weights Stable Diffusion family (1.x, SDXL, SD3) under CreativeML Open RAIL-M licences that enabled the global open-source community ecosystem. AstraZeneca (Cambridge) and GSK (London) are active users of RFdiffusion and molecular diffusion tools in their computational drug discovery pipelines. Framestore and Moving Picture Company (London VFX studios) are early adopters of diffusion-based AI for visual effects and post-production workflows.

  The UK government’s AI Action Plan (January 2025) explicitly mentions generative AI as a priority investment area. The Creative Industries sector (which accounts for ~£115 billion GVA) has engaged with the Intellectual Property Office on guidance regarding copyright status of AI-generated images, a debate that is particularly salient given that Stability AI was subject to lawsuits from Getty Images (UK High Court proceedings initiated 2023) regarding training data usage.

  Future Directions (2026–2030)

  The most active research frontiers include: (1) World models — extending video diffusion to persistent, physically-consistent world representations that can be interacted with, bridging diffusion models and reinforcement learning environments; (2) Multimodal joint generation — unified diffusion models that generate text, image, audio, and video from a single latent space, advancing the Multimodal Model paradigm; (3) Scientific generative modelling — diffusion over molecular dynamics trajectories, crystal structure generation, and climate simulation emulation; (4) Diffusion policy in robotics — using diffusion models as action sequence generators in robot learning, with Diffusion Policy (Chi et al. 2023) demonstrating visuomotor policy learning from demonstrations; (5) Personalisation and compositionality — efficient fine-tuning methods (DreamBooth, LoRA) enabling user-personalised generation at inference time; (6) Real-time diffusion — continued distillation and caching research pushing latency below 1 second for interactive creative applications; (7) Privacy-preserving generation — combining diffusion models with Differential Privacy to generate formally private Synthetic Data for healthcare and finance; (8) 3D and 4D generation — extending DiT architectures to 3D Gaussian Splat and NeRF representations for Spatial Computing and Digital Twin content creation.

  Research & Literature

  1. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., & Ganguli, S. (2015). “Deep Unsupervised Learning using Nonequilibrium Thermodynamics.” ICML 2015, PMLR.
  2. Hyvärinen, A. (2005). “Estimation of Non-Normalized Statistical Models by Score Matching.” Journal of Machine Learning Research, 6, 695–709.
  3. Song, Y., & Ermon, S. (2019). “Generative Modeling by Estimating Gradients of the Data Distribution.” NeurIPS 2019.
  4. Ho, J., Jain, A., & Abbeel, P. (2020). “Denoising Diffusion Probabilistic Models.” NeurIPS 2020. (DDPM — foundational formulation.)
  5. Song, J., Meng, C., & Ermon, S. (2021). “Denoising Diffusion Implicit Models.” ICLR 2021. (DDIM — deterministic fast sampling.)
  6. Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., & Poole, B. (2021). “Score-Based Generative Modeling through Stochastic Differential Equations.” ICLR 2021. (SDE unification — VP-SDE, VE-SDE.)
  7. Dhariwal, P., & Nichol, A. (2021). “Diffusion Models Beat GANs on Image Synthesis.” NeurIPS 2021. (Classifier guidance; demonstrated diffusion superiority over GANs.)
  8. Nichol, A., & Dhariwal, P. (2021). “Improved Denoising Diffusion Probabilistic Models.” ICML 2021. (Cosine noise schedule; learnable variance.)
  9. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., & Ommer, B. (2022). “High-Resolution Image Synthesis with Latent Diffusion Models.” CVPR 2022. (LDM — Stable Diffusion basis.)
  10. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., & Chen, M. (2022). “Hierarchical Text-Conditional Image Generation with CLIP Latents.” arXiv:2204.06125. (DALL-E 2.)
  11. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., … & Norouzi, M. (2022). “Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding.” NeurIPS 2022. (Imagen.)
  12. Ho, J., & Salimans, T. (2022). “Classifier-Free Diffusion Guidance.” NeurIPS 2022 Workshop. (CFG — the standard conditioning mechanism.)
  13. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., & Zhu, J. (2022). “DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in around 10 Steps.” NeurIPS 2022.
  14. Peebles, W., & Xie, S. (2023). “Scalable Diffusion Models with Transformers.” ICCV 2023. (DiT architecture.)
  15. Liu, X., Gong, C., & Liu, Q. (2022). “Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow.” ICLR 2023. (Rectified Flow.)
  16. Lipman, Y., Chen, R. T. Q., Ben-Hamu, H., Nickel, M., & Le, M. (2023). “Flow Matching for Generative Modeling.” ICLR 2023.
  17. Watson, J. L., Juergens, D., Bennett, N. R., et al. (2023). “De novo design of protein structure and function with RFdiffusion.” Science, 380(6647), 1123–1131. (RFdiffusion protein design.)
  18. Esser, P., Kulal, S., Blattmann, A., et al. (2024). “Scaling Rectified Flow Transformers for High-Resolution Image Synthesis.” ICML 2024. (Stable Diffusion 3 / MMDiT.)
  19. Chi, C., Feng, S., Du, Y., et al. (2023). “Diffusion Policy: Visuomotor Policy Learning via Action Diffusion.” RSS 2023. (Robotics diffusion policy.)
  20. Black Forest Labs (2024). “FLUX.1 Technical Report.” Black Forest Labs. (FLUX architecture; basis of MLPerf Training 5.0 benchmark.)
  21. MLCommons (2025). “MLPerf Training Introduces Flux.1 Text-to-Image Benchmark.” https://mlcommons.org/2025/10/training-flux1/
  22. Nishimura, J. (2025). “The Architecture Behind Every State-of-the-Art Image Generator: DiT + Rectified Flow, Explained.” Medium / ICLR Blogposts 2026. https://iclr-blogposts.github.io/2026/blog/2026/diffusion-architecture-evolution/
  23. DiffMC-Gen Authors (2025). “DiffMC-Gen: A Dual Denoising Diffusion Model for Multi-Conditional Molecular Generation.” PMC. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12165109/
  24. Avrahami, Z., et al. (2025). “Diffusion Models at the Drug Discovery Frontier: A Review on Generating Small Molecules Versus Therapeutic Peptides.” Biology, 14(12), 1665. https://doi.org/10.3390/biology14121665
  25. Efficient Diffusion Models Authors (2024). “Efficient Diffusion Models: A Comprehensive Survey from Principles to Practices.” arXiv:2410.11795.
  26. C2PA (2024). “C2PA Specification v2.0 — Coalition for Content Provenance and Authenticity.” https://c2pa.org/specifications/specifications/
  27. GenRA / AIUnpacking (2026). “DALL-E Is Dead: OpenAI Retires Its Image Models on May 12.” https://aiunpacking.com/review/dall-e/; https://genra.ai/blog/dall-e-retired-may-2026-what-replaces-it
  28. Kittl Blog (2026). “AI image generation complete guide for designers in 2026.” https://www.kittl.com/blogs/ai-image-generation-guide-ais/

Provenance