Diffusion Model
A diffusion model is a probabilistic generative machine learning model architecture whose core principle involves gradually adding noise to data (forward process) and then learning to progressively denoise and reconstruct the data (reverse process) to model the data distribution. Originally applied extensively to image generation, diffusion models have since expanded to generative tasks across modalities including audio, video, and text, and are characterized by high generation quality and strong diversity.
Core Facts
Diffusion models circumvent the instability of adversarial training through a Markov chain mechanism that progressively adds Gaussian noise and then trains the model to learn the reverse denoising process.
扩散模型的核心原理分为前向扩散过程（逐步添加高斯噪声）和反向去噪过程（从噪声中还原图像）两个阶段
DDPM (Denoising Diffusion Probabilistic Models) laid the theoretical foundation for diffusion models in 2020.
Timeline (last 90 days)
The computational complexity of diffusion models in the temporal dimension grows approximately quadratically with duration, and memory and compute consumption rise sharply.
Current mainstream text-to-image and image-to-image tools have lowered the barrier for 'generating a photo of someone wearing specific clothing' to just a text description or a reference image, taking seconds and costing almost nothing.
Methods based on diffusion models and large-scale vision-language models can already infer geometric structure using a few photos taken with an ordinary phone, no longer requiring professional photogrammetry software or precise camera calibration.
Latent Space Constraint, by imposing specific constraints on the latent space during the inference stage, can greatly reduce the variance of generated results without retraining the model, trading generation diversity for speed and stability in specific scenarios.
Common methods for photo-to-video animation include deformation driven by facial keypoints, three-dimensional reconstruction based on neural radiance fields (NeRF), and direct generation of temporal frames by diffusion models.
In diffusion models, the score function is used to describe the log gradient of the data distribution and is trained via denoising score matching.
The diffusion model's pixel-by-pixel, probability-distribution-based generation method determines that it lacks character-level precise control capability, learning the visual statistical features of text rather than character-by-character typesetting.
Diffusion models are prone to text hallucination when generating short text, manifesting as extra strokes, confusion of similar letters (such as S and 5, O and 0), or inconsistent font weight and letter spacing.
Text-to-video tools fall into roughly two categories: the text-plus-stock-library matching model represented by Fliki offers higher controllability, while the diffusion model/GAN-generated original footage model represented by Zebracat offers stronger visual uniqueness but lower stability.
Traditional image generation or prediction models (such as diffusion models and discriminative classifiers) typically perform single-pass forward inference and have no built-in self-correction mechanism.
40 more timeline events
All Facts (20)
Diffusion models circumvent the instability of adversarial training through a Markov chain mechanism that progressively adds Gaussian noise and then trains the model to learn the reverse denoising process.
90%Verified扩散模型的核心原理分为前向扩散过程（逐步添加高斯噪声）和反向去噪过程（从噪声中还原图像）两个阶段
90%VerifiedDDPM (Denoising Diffusion Probabilistic Models) laid the theoretical foundation for diffusion models in 2020.
90%VerifiedStable Diffusion is based on the Latent Diffusion Model and generates high-quality images by performing a denoising process in a compressed latent space
90%VerifiedCompared to GANs, diffusion models are more stable to train, generate greater diversity, and are less prone to mode collapse.
90%VerifiedDiffusion models are inspired by non-equilibrium thermodynamics and employ a two-step training process: the forward process gradually adds Gaussian noise to an image until it becomes pure random noise, and the reverse process trains a neural network to progressively denoise from pure noise back to a clear image.
80%VerifiedCompared to GANs, diffusion models offer more stable training, higher generation diversity, and easier integration with text conditioning.
80%VerifiedText instructions are converted into semantic vectors through cross-modal encoders such as CLIP, embedding text and images into the same semantic space to guide image generation.
80%VerifiedGANs improve generation quality through adversarial interplay between a generator and a discriminator, but are prone to mode collapse; diffusion models have more stable training objectives and naturally avoid mode collapse.
80%VerifiedThe core idea of diffusion models originates from non-equilibrium thermodynamics: during the training phase, Gaussian noise is gradually added to clear images, and during the inference phase, noise is progressively removed from random noise in reverse to restore the image.
80%VerifiedCurrent mainstream text-to-video models are generally based on diffusion model architectures, generating coherent video frame sequences by extending image generation capability along the temporal dimension.
80%VerifiedThe output randomness of image generation models fundamentally stems from the initial noise sampling in the diffusion process, with the Seed (random number seed) controlling the initial noise.
80%VerifiedThe theoretical foundations of diffusion models can be traced back to a 2015 paper on non-equilibrium thermodynamics published by Sohl-Dickstein et al.
80%VerifiedThe denoising process of diffusion models is based on the DDPM principle, which involves progressively adding Gaussian noise to data during training and then training a neural network to learn the reverse denoising process.
80%VerifiedThe results of AI image expansion differ each time, and this randomness stems from the diffusion model starting from different random noise each time
80%VerifiedMost mainstream AI image generation models are currently based on the Diffusion Model architecture.
80%VerifiedAI video generation technology has undergone a paradigm shift from GANs (Generative Adversarial Networks) to Diffusion Models.
80%Verified扩散模型通过在海量图像数据上训练，学会了从随机噪声中逐步去噪生成图像的能力
80%VerifiedQwen's image generation capability is based on Diffusion Model technology.
80%VerifiedTraditional video generation relies on diffusion model technology, generating images or video frames from random noise through a step-by-step denoising process, with each frame requiring tens to hundreds of iterative computation steps.
75%