Amazing Generative Models

Walk into Generative Models

1. What are Diffusion Models
   • Supplementary papers
   1. Denoising Diffusion Probabilistic Models
   2. Denoising Diffusion Implicit Models
   3. High-Resolution Image Synthesis with Latent Diffusion Models
2. An Introduction to Flow Matching
   • Supplementary papers
   1. Flow Matching for Generative Modeling
   2. Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

Works

Principle

Mean Flows for One-step Generative Modeling (NeurIPS 2025) [Paper] [Code]

• Compute straight marginal flow from conditional flow to achieve one-step inference.

Diffusion Transformers with Representation Autoencoders [Homepage] [Paper] [Code]

• New form of image encoder for generative model, leverage strong semantic representation from pretrained foundation image encoder like DINOv2.

On the Edge of Memorization in Diffusion Models [Paper]

• When do diffusion models generate and when do diffusion models memorize?

Applications

Dynamic Typography: Bringing Text to Life via Video Diffusion Prior (ICCV 2025) [Homepage] [Paper] [Code] [Demo]

• Generate a dynamic typography based on video foundation model, optimized by SDS.

3D Generation & Representation

SparseFlex: High-Resolution and Arbitrary-Topology 3D Shape Modeling (ICCV 2025) [Homepage] [Paper] [Code]

• The SOTA work for sparse-structured (or VAE-based) latent representation.

Conditional Generation

LOTS of Fashion! Multi-Conditioning for Image Generation via Sketch-Text Pairing (ICCV 2025) [Homepage] [Paper] [Code]

• They build the condition with multiple image layers of sketch and text condition, and keep them as seperate conditions jointly control the image generation. (Maybe this work just for garments? but it can inspire other layer editing)

Editing

FlowEdit: Inversion-Free Text-Based Editing Using Pre-Trained Flow Models (ICCV 2025) [Homepage] [Paper] [Code]

• Inversion-free generative editing, modeling the editing flow on data manifold, avoiding inversing the image to noise.

Video in 3D

ReCamMaster: Camera-Controlled Generative Rendering from A Single Video (ICCV 2025) [Homepage] [Paper] [Code]

• Generating video in novel camera trajactory from reference video clips. (Btw, the clips in demo are nice)

Generating in Other Ways

Discrete Distribution Network (ICLR 2025) [Homepage] [Paper] [Code] [Demo]

• Generating results by selecting candidate from cascade hierarchy network layers.

Tools

Conditioned Generation

1. ControlNet++ A novel ControlNet Architecture and also a good practice, the repo provide 10+ pre-trained conditioned generative model.

Development