ACE++ - AI Image Tools Tool

Overview

ACE++ (ACE_Plus) is an instruction-based image creation and editing toolkit developed by Tongyi Lab / Alibaba (ali-vilab). It unifies reference-driven generation, subject-consistent generation, and local (mask-based) editing in a single diffusion-based framework that uses context-aware content filling and a Long-context Condition Unit to accept long instruction sequences and multi-image inputs. The project publishes lightweight LoRA variants (portrait, subject, local editing) and a fully-finetuned FFT model; the authors recommend LoRA for stable, vertical workflows while providing the FFT model for broader research use. ([huggingface.co](https://huggingface.co/ali-vilab/ACE_Plus)) ACE++ was released to the community in early 2025 with code, demos, and model checkpoints on Hugging Face and GitHub. The repository includes a Gradio demo, ComfyUI workflow examples, and training/inference scripts (LoRA training scripts and an FFT config). Community builders have produced ComfyUI workflows, video tutorials, and guides demonstrating face-swap, logo-paste, try-on, and regional edit applications; at the same time users and the authors note artifacts and some instability on complex tasks. ([huggingface.co](https://huggingface.co/ali-vilab/ACE_Plus))

Model Statistics

• Downloads: 126
• Likes: 299

Model Details

Architecture and training: ACE++ is an instruction-based diffusion framework that extends ACE’s Long-context Condition Unit (LCU) and uses a two-stage training scheme: stage one leverages 0-ref tasks to adapt a text-to-image foundation (the authors use FLUX.1-Fill-dev as the base), and stage two finetunes on the full set of ACE tasks (repainting, regional editing, subject/reference generation, etc.). The FFT (fully finetuned) variant introduces an additional 64-channel extension to the FLUX Fill input to separate repainting and editing signals (changing input channels from 384 to 448). The paper and code describe the LCU, two-stage training, and instruction format used to combine reference images, edit images, and edit masks into a single conditioning input. ([arxiv.org](https://arxiv.org/abs/2501.02487)) Released artifacts and recommended usage: the project provides three LoRA models (portrait, subject, local editing — commonly packaged as comfyui_portrait_lora64.safetensors, comfyui_subject_lora16.safetensors, comfyui_local_lora16.safetensors) and an FFT model. The authors recommend LoRA models for better stability and vertical use cases, while the FFT model is intended for research and broader image-to-image tasks. The repo includes inference scripts (infer_lora.py / infer_fft.py), demo scripts (demo.py, demo_lora.py, demo_fft.py), and ComfyUI workflow examples; max_seq_length (2048–8192) is exposed to control memory/clarity tradeoffs. Known limitations include artifacts (hands, some object add/remove failures) and instability when fully post-training on FLUX. ([github.com](https://github.com/ali-vilab/ACE_plus))

Key Features

• Instruction-based image generation and editing using natural-language prompts.
• Portrait-consistent generation to preserve facial identity across scenes.
• Subject-driven generation to keep object/logo/model consistency across contexts.
• Local (mask-guided) regional editing and repainting for precise area changes.
• LoRA checkpoints for lightweight fine-tuning (portrait/subject/local variants).
• FFT model with added 64 channels to separate repainting and editing inputs.
• ComfyUI workflows and a Gradio demo for rapid prototyping and community use.

Example Usage

Example (python):

import os
import subprocess

# Example: run the provided Gradio demo or LoRA demo as described in the repository README.
# Set model paths (can point to Hugging Face scepter paths or local model files).
os.environ["FLUX_FILL_PATH"] = "hf://black-forest-labs/FLUX.1-Fill-dev"
os.environ["PORTRAIT_MODEL_PATH"] = "hf://ali-vilab/ACE_Plus@portrait/comfyui_portrait_lora64.safetensors"
os.environ["SUBJECT_MODEL_PATH"] = "hf://ali-vilab/ACE_Plus@subject/comfyui_subject_lora16.safetensors"
os.environ["LOCAL_MODEL_PATH"] = "hf://ali-vilab/ACE_Plus@local_editing/comfyui_local_lora16.safetensors"

# Launch the demo (run in the checked-out ACE_plus repository directory)
subprocess.run(["python", "demo.py"])

Benchmarks

Hugging Face likes: 299 likes (model card) (Source: https://huggingface.co/ali-vilab/ACE_Plus)

Hugging Face downloads (last month): 126 downloads (last month) (Source: https://huggingface.co/ali-vilab/ACE_Plus)

GitHub stars: ~1.3k stars (Source: https://github.com/ali-vilab/ACE_plus)

ArXiv (paper): ACE++ arXiv:2501.02487 (v3 last revised Jan 15, 2025) (Source: https://arxiv.org/abs/2501.02487)

Key Information

• Category: Image Tools
• Type: AI Image Tools Tool