Diffusion Model
Core Facts
Compared to GANs, diffusion models are more stable to train, generate greater diversity, and are less prone to mode collapse.
Diffusion 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.
Compared to GANs, diffusion models offer more stable training, higher generation diversity, and easier integration with text conditioning.
Timeline (last 90 days)
The maturation of diffusion models has greatly improved the naturalness of virtual try-on image compositing, enabling simulation of lighting, wrinkles, and occlusion relationships in clothing.
The advantage of the code generation path over diffusion models is that the output is cleaner and editable, and VRAM consumption is concentrated mainly in the language model inference stage rather than the image generation stage.
Unifying generation and editing into a single set of weights typically follows two technical approaches: a conditional generation framework based on diffusion models, and a paradigm drawing on instruction fine-tuning.
The T2V technical approach commonly employs a combination of diffusion models and Transformer architectures.
The article concludes that the corruption caused by the seed stems from numerical conflicts in the sampling process itself and is unrelated to training data.
In image editing tasks, the model adds a certain amount of noise to the input image before denoising, rather than starting from pure noise.
Diffusion models require tens or even hundreds of forward inference passes to generate each frame
Mapping facial features onto a cat image technically relies on image style transfer or diffusion-model-based conditional generation methods.
Diffusion model approaches can handle diverse inputs more flexibly than earlier fixed filters, but are more sensitive to input photo quality and front-facing angles.
The transition problem in AI video stems from the autoregressive generation approach of models, where each video clip is independently sampled and generated, lacking cross-clip temporal consistency constraints.
24 more timeline events
All Facts (20)
Compared 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%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 video diffusion model adopts a Diffusion Transformer (DiT) architecture and models both spatial and temporal dimensions simultaneously by introducing a 3D spatiotemporal attention mechanism in the temporal attention layers.
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%VerifiedJimeng is an AI image and video generation platform launched by ByteDance, whose image generation capability is based on Diffusion Model technology.
80%VerifiedQwen's image generation capability is based on Diffusion Model technology.
80%VerifiedEvery user query to a large AI model company requires substantial GPU compute for inference, which is a variable cost that grows linearly with usage rather than a fixed cost.
75%VerifiedThe transition problem in AI video stems from the autoregressive generation approach of models, where each video clip is independently sampled and generated, lacking cross-clip temporal consistency constraints.
65%VerifiedThe multi-stage Cascaded Diffusion Model architecture first generates the overall composition at low resolution, then progressively upscales it through super-resolution networks to supplement high-frequency details.
65%VerifiedScalar motion transfer generates character dynamics by referencing existing motion videos, a principle that typically involves skeletal pose sequence extraction.
65%Verified扩散模型在推理阶段从随机高斯噪声出发，经数十至数百步迭代去噪得到输出，即便固定随机种子，提示词的细微差异也会导致人脸特征、服装细节的显著偏移
65%VerifiedThe core idea of the video-to-video guided generation paradigm originates from the conditional generation mechanism of diffusion models, which introduces additional visual conditioning signals during the denoising process to constrain the generation direction.
65%Verified扩散模型在生成图像时本质上是在高维潜空间（Latent Space）中进行去噪采样，缺乏多角度约束时同一角色在不同帧中的特征易产生语义漂移
65%VerifiedAI 图像生成中通过视觉编码器（如 CLIP 模型）对输入产品照片进行特征提取，再结合文本提示词通过扩散模型逐步去噪生成目标图像
65%Unverified现代角色动作迁移技术基于扩散模型，通过ControlNet等条件控制机制将姿态骨骼图作为额外条件输入，利用IP-Adapter等技术保持外观一致性
90%