2025 년 이미지 생성모델을 위한 주요 모델 리서치

study 2025-

Relightful Harmonization: Lighting-aware Portrait Background Replacement
https://openaccess.thecvf.com/content/CVPR2025/papers/Wang_Towards_Enhanced_Image_Inpainting_Mitigating_Unwanted_Object_Insertion_and_Preserving_CVPR_2025_paper.pdf
https://arxiv.org/pdf/2506.14549 - dreamlight
https://openaccess.thecvf.com/content/CVPR2025/papers/Wang_Comprehensive_Relighting_Generalizable_and_Consistent_Monocular_Human_Relighting_and_Harmonization_CVPR_2025_paper.pdf

https://arxiv.org/pdf/2506.18898

• Video generation
  Consistent 3D Camera Motion
• diffusion edit (edit 보다는 컨셉유지)
  Diffusion Self-Distillation for Zero-Shot Customized Image Generation : zeroshot editing
  OminiControl / FLUX -> anystory.
  ic-custom
  Vision as a Dialect
  RealCustom++ (subject customization
  https://github.com/River-Zhang/ICEdit
• customization & instruction (preservation)
  dreamO : https://arxiv.org/pdf/2504.16915
  ace++: Instruction-Based Image Creation and Editing via Context-Aware Content Filling
  RealCustom++ (subject customization
  infiniteyou (https://arxiv.org/pdf/2503.16418 )
  https://github.com/shhh0620/IDS
  PULID Pure and Lightning ID Customization via Contrastive Alignment
  magicfusion identity 를 text encoding 쪽에 학습
  간단한 Styling 도되는듯
• prompt enhancer
  https://huggingface.co/roborovski/superprompt-v1
  https://huggingface.co/gokaygokay/Lamini-Prompt-Enchance-Long
• object remover
  https://github.com/Yikai-Wang/asuka-misato https://arxiv.org/pdf/2312.04831
• language model
  https://arxiv.org/pdf/2506.18898
  WorldScribe: Towards Context-Aware Live Visual Descriptions
• concept eraser
  Concept Pinpoint Eraser for Text-to-image Diffusion Models via Residual Attention Gate (ICLR 2025)
• matrial transfer
  https://ttchengab.github.io/zest/
  NeRF - Texture : Texture Synthesis with Neural Radiance Fields (2023,SIGGRAPH)
  https://github.com/astra-vision/MaterialPalette : PBR map 까진만들어짐, Generation 은 top view 만됨
  https://github.com/astra-vision/MatSwap :
  TextureDreamer: Image-guided Texture Synthesis through ...
  FabricDiffusion: High-Fidelity Texture Transfer for 3D ...
  https://github.com/ControlGenAI/MaterialFusion
• recolor?
  https://enis.dev/uip2p/ instructpix2pix 거꾸로해서 Recon 하는 로스 조금더 안정적인느낌
  https://github.com/deepffff/SADis
  부모 InstantStyle
• MLLM
  https://github.com/mc-lan/Awesome-MLLM-Segmentation?tab=readme-ov-file
  https://github.com/BradyFU/Awesome-Multimodal-Large-Language-Models
• inpaint
  smart brush : Foreground mask 를 같이 diffusion model 에 넣고 학습하는 것이 특징
  Foreground mask 만 디노이즈함
• fashion layered diffusion
  https://intelligolabs.github.io/lots/
• inpainting 꽌련
  -- https://github.com/yayoyo66/RDIStitcher
  : 2개의 다른 영역을 합성할 때 Stich 하는게 특징
  -- Erase Diffusion: Empowering Object Removal Through Calibrating
  Diffusion Pathways : erase inpainting 을 위한 finetuned model, code 공개 x ERASE될 OBJECT 를 서서히 SHRINK 하도록함. (다른물체가 생기는걸 방지 하기 위해서서)

Identity-preserving Distillation Sampling by Fixed-Point Iterator
Latent space Super-Resolution for Higher-Resolution Image Generation with Diffusion Models
Prompt-Driven Contrastive Learning for Transferable Adversarial Attacks
Random Conditioning for Diffusion Model Compression with Distillation

Universal Image Generation and Editing via Learning Real-world Dynamics : flux-kontext, qwen edit model 에 가까운구조