Alibaba Open Sources ACE++: Zero-Training Character-Consistent Image Generation

news

Alibaba Research Institute open sources image generation tool ACE++, supporting character-consistent image generation from single input through context-aware content filling technology, offering online experience and three specialized models.

February 10, 2025 — Alibaba Research Institute officially announces the open-source release of next-generation AI imaging tool ACE++. Based on innovative context-aware content filling algorithms, users can generate new images with highly consistent character features from just a single input image, supporting both online experience and local deployment.

Core Technical Innovations

ACE++ Architecture Diagram

Key Features

  • Zero-Training Generation: Leveraging FLUX.1-Fill-dev base model, achieving training-free deployment through LoRA adaptation
  • Multimodal Editing:
    • Character Outfit Change (supports clothing/hairstyle/accessory changes)
    • Scene Reconstruction (background replacement/object addition/removal)
    • Smart Restoration (flaw removal/quality enhancement)
  • Semantic Understanding: Can parse compound instructions like "add steam to coffee cup, place on wooden table"

Technical Breakthroughs

  • Long Context Unit (LCU): Simultaneously processes image content, text instructions, and editing regions
  • Dynamic Attention Mechanism: Achieves 92.3% feature retention rate at 512×512 resolution
  • Two-Stage Optimization: Step-by-step training approach combining basic restoration and specialized editing skills

Application Scenarios Testing

| ACE_plus | ACE_plus | |

-|

-| | ACE_plus | ACE_plus |

Typical Applications

  1. Virtual Model Outfit Changes

    • Generate multi-angle displays from flat clothing images
    • Supports dynamic adjustment of skin tone/body type/scene

  2. Film Character Design

    • Achieves cross-style transformation (realistic→Disney/cyberpunk)
    • Multi-scene generation maintaining character feature continuity

  3. Smart Image Restoration

    • 4K resolution reconstruction of old photos
    • Seamless removal of complex occlusions

Resource Access Channels

Official Entries

| Resource Type | Access Link | |

--|

-| | Project Homepage | https://ali-vilab.github.io/ACE_plus_page/ | | Code Repository | GitHub | | Online Experience | ModelScope |

Specialized Adaptation Models

| Model Type | File Name | ModelScope Download | HuggingFace Download | |

|

--|

-|

| | Portrait Generation | comfyui_portrait_lora64.safetensors | Portrait Model | Portrait Model | | Object Transfer | comfyui_subject_lora16.safetensors | Subject Model | Subject Model | | Local Editing | comfyui_local_lora16.safetensors | LocalEditing Model | LocalEditing Model |

Base Dependency Models

| Model Name | Download Channel | |

|

--| | FLUX.1-Fill-dev | HuggingFace Download | | Flux-Fill FP8 | CivitAI Download |

Technical Development Outlook

The current version still has room for improvement in complex object processing (hand detail accuracy 62.3%) and Chinese text support. The development team reveals plans to launch video continuous frame editing functionality in Q3 2025, and will release the complete ACE++ Fully model by year-end.

Please wait for workflow files after ComfyUI Wiki testing for content updates View Workflow Updates

Alibaba Open Sources ACE++: Zero-Training Character-Consistent Image Generation | ComfyUI Wiki