The Alibaba team previously released ACE++, featuring three LoRA models based on FLUX, which provided exceptional consistency in transferring objects and people. Now, they have introduced a major upgrade: The Unified FFT Model.

ACE_Plus: Introduction to the Unified FFT Model

• GitHub Repository: https://github.com/ali-vilab/ACE_plus

• Model Download: https://hf-mirror.com/ali-vilab/ACE_Plus/tree/main

  

The Final ACE Iteration – A New Direction

The team has stated that this will likely be the final iteration of the ACE series. The FFT Unified Model is now capable of handling image generation, local editing, and controlled generation effectively, eliminating the need for further updates.
The main reason for this decision is the high heterogeneity between training datasets and the FLUX model, which makes training extremely unstable. Moving forward, the team will focus on WAN 2.1 to explore new research directions.

ACE FFT Unified Model: Enhanced Image Processing Capabilities

The model is fully fine-tuned using ACE data and supports various editing and reference-based generation tasks, including:

• Repainting

• Outline Redraw

• Depth Redraw

• Recoloring

• Region Editing

• Super Resolution

  

ComfyUI Integration & Installation

The official release includes ComfyUI node packages, workflows, and the FFT model.
CloserAI members can access the ACE++ FFT model, workflows, and nodes directly from the model library.

ComfyUI Installation Steps

1. Copy the workflow/ComfyU-ACE_Plus folder into the ComfyUI custom_nodes directory.

2. Launch ComfyUI and find four example workflows in the workflow_example_fft folder:

   • workflow_no_preprocess.json – Uses preprocessed images (e.g., depth maps, outlines) as input or for super-resolution tasks.

   • workflow_controlpreprocess.json – Enables controlled image-to-image transformations.

   • workflow_reference_generation.json – Supports reference-based image generation for portraits or objects.

   • workflow_referenceediting_generation.json – Allows reference-based image editing.

Optimize GPU Memory Usage

The ComfyUI version includes a max_seq_length parameter to adjust GPU memory consumption during inference.

• Range: 1024 - 5120

• Higher values result in clearer images but require more VRAM.

• Lower values reduce memory usage but may affect image clarity.

The ACE++ FFT Unified Model unlocks next-level image processing—download and try it today! 🚀