ACE++ (Alibaba / Tongyi Lab)¶

Unified multi-task image editing model built on [[FLUX Kontext|FLUX.1-Fill-dev]]. Supports try-on, face swap, inpainting, style transfer, super-resolution, logo placement — all through LoRA adapters or full fine-tune.

Paper: arXiv:2501.02487. GitHub: ali-vilab/ACE_plus.

Architecture¶

LCU++ (Long-context Condition Unit++)¶

Key change from original ACE: channel concatenation instead of sequence concatenation.

Original LCU:  [cond_image; mask] + [noise; mask]  → sequence concat (2× attention cost)
LCU++:         [cond_image; mask; noise]            → channel concat (efficient)

For N-ref tasks: multiple references concat along sequence, but each unit uses channel concat internally.

FFT model specifics: - IN_CHANNELS: 448 (vs 384 for FLUX.1-Fill-dev) — 64 extra for edit image latent - Same transformer: 3072 hidden, 24 heads, 19+38 blocks - REDUX_DIM: 1152 for image embedding

Model Variants¶

3 LoRA models (recommended):

1 FFT model (fully fine-tuned): all tasks but less stable. Authors explicitly recommend LoRA over FFT.

Training¶

Two-stage: 1. Pre-train on 0-ref tasks (T2I, inpainting) starting from FLUX.1-Fill-dev 2. Fine-tune on all tasks (0-ref + N-ref) from ACE dataset

Critical Warning¶

Development on FLUX base is SUSPENDED. Authors found "high degree of heterogeneity between training data and FLUX model" causing highly unstable training. Future ACE versions will use Alibaba's Wan foundation models instead.

License¶

Inherits FLUX.1-dev Non-Commercial License — non-commercial only without BFL licensing.

Gotchas¶

• FFT model quality worse than LoRA models — use LoRA variants
• FLUX.1-dev is guidance-distilled → negative prompts have uncertain effect
• Requires FLUX.1-Fill-dev as base (not plain FLUX.1-dev)
• No quantitative benchmarks published — qualitative results only

Key Links¶

• GitHub: github.com/ali-vilab/ACE_plus
• HF: huggingface.co/ali-vilab/ACE_Plus
• Demo: huggingface.co/spaces/scepter-studio/ACE-Plus