Research map / Image generation / Editing and style transfer
FluxSpace: Disentangled Semantic Editing in Rectified Flow Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FluxSpace, a training-free inference-time image editing method that edits the joint attention outputs and pooled CLIP embeddings of the Flux rectified flow transformer to get disentangled fine-grained and coarse semantic edits.
Problem
Rectified flow transformers such as Flux struggle with disentangled, attribute-specific image editing; existing editing methods either entangle unrelated attributes or require per-edit training.
Approach
Treats the outputs of Flux's joint attention layers as a linear representation space: orthogonally projects the edit-conditioned attention output against the null-prompt attention output to get a semantic direction, adds it with a scale, and masks it with a text-attention map; coarse edits are applied by projecting the pooled CLIP embedding in the same way.
Key results
- On the Smile edit, FluxSpace reaches CLIP-I 0.9417 and DINO 0.9402 on eyeglasses, the best content preservation among compared methods.
- User study with 50 participants gives FluxSpace an overall preference score of 4.19 out of 5, versus 3.34 for the next best (Sliders-FLUX).
- Method runs in about 20 seconds per image on a single NVIDIA L40 GPU with 30 generation steps.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FluxSpace attention editing instead of competing editing methods | FLUX.1-dev image editing · Eyeglasses and Smile | CLIP-I and DINO: Eyeglasses CLIP-I: 0.9354 -> 0.9417; DINO: 0.9359 -> 0.9402; Smile CLIP-I: 0.8834 -> 0.9038; DINO: 0.9134 -> 0.9347 | Helped |
| Use FluxSpace attention editing instead of competing editing methods | FLUX.1-dev image editing · Eyeglasses and Smile | CLIP-T: Eyeglasses: 0.2878 -> 0.3180; Smile: 0.3438 -> 0.3503 | Mixed |
| Use FluxSpace instead of competing methods | Image editing · Overall | User Pref.: 3.34 -> 4.19 | Helped |
| Add pooled-text coarse editing with λcoarse (ablation) | FluxSpace · Comics-style editing | Style control | Helped |
| Vary the fine-grained edit scale λfine (ablation) | FluxSpace · Gender and Smile edits | Edit strength | Helped |
| Add attention masking using threshold τm (ablation) | FluxSpace · Sunglasses editing | Disentanglement | Helped |
| Start editing at earlier denoising timesteps (ablation) | FluxSpace · Image editing | Edit strength and consistency | Mixed |
Surprising findings
- TurboEdit achieves higher CLIP-T text alignment scores but substantially changes subject identity, so text alignment alone is misleading for editing quality.
Limitations
- Potential misuse for deceptive or non-consensual face manipulation, as the authors acknowledge.
- Real-image editing relies on RF-Inversion, whose inversion does not fully map the input into the generator's latent distribution, so the authors focus on generated images.
Experimental setup
Benchmarks: 60-image face editing set (eyeglasses and smile edits)
Metrics: CLIP-T CLIP-I DINO user preference (Likert 1-5)
Baselines: LEDITS++ TurboEdit Sliders-FLUX RF-Inversion
Builds on: Flux RF-Inversion MM-DiT linear representation hypothesis
Closest papers by meaning
- SplitFlow: Flow Decomposition for Inversion-Free Text-to-Image Editing (NeurIPS 2025 · similarity 0.89)
- Latent Space Editing in Transformer-Based Flow Matching (AAAI 2024 · similarity 0.87)
- Semantic Image Inversion and Editing using Rectified Stochastic Differential Equations (ICLR 2025 · similarity 0.87)
- Improving Editability in Image Generation with Layer-wise Memory (CVPR 2025 · similarity 0.85)
- Learning to Manipulate Artistic Images (AAAI 2024 · similarity 0.85)
- Scene Graph Disentanglement and Composition for Generalizable Complex Image Generation (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.