Research map / Image generation / Editing and style transfer
Tuning-Free Inversion-Enhanced Control for Consistent Image Editing
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 Tuning-free Inversion-enhanced Control (TIC), which injects DDIM-inversion self-attention keys/values into the sampling process to improve reconstruction and enable non-rigid, content-consistent real-image editing without fine-tuning.
Problem
Consistent editing of real images requires non-rigid edits while preserving identity and attributes, but fine-tuning methods are slow and tuning-free methods are limited by poor DDIM inversion reconstruction.
Approach
Replace the key and value features in self-attention layers during DDIM sampling with those from the inversion trajectory (at the previous timestep), reducing the noise-prediction mismatch that causes reconstruction error; extend with mask-guided attention concatenation of inversion and DDIM-edit keys/values, and combine with ControlNet.
Key results
- TIC reconstruction reaches PSNR 27.11 and SSIM 0.7864, close to the VAE upper bound (27.17 / 0.7886) and above DDIM (22.95 / 0.6840).
- TIC runs in 5.56 s per image, matching DDIM and far faster than NTI (97.86 s at 250 iterations) and PTI (89.89 s at 250 iterations).
- Qualitatively, TIC performs non-rigid edits (e.g., raising a hand) while keeping background content, where MasaCtrl introduces background artifacts.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace sampling self-attention keys and values with inversion features | TIC reconstruction · MS-COCO 2017 validation | PSNR: 22.95 -> 27.11 | Helped |
| Replace sampling self-attention keys and values with inversion features | TIC reconstruction · MS-COCO 2017 validation | SSIM: 0.6840 -> 0.7864 | Helped |
| Use TIC instead of NTI with 250 iterations | Image reconstruction · MS-COCO 2017 validation | PSNR: 25.70 -> 27.11 | Helped |
| Use TIC instead of NTI with 500 iterations | Image reconstruction · MS-COCO 2017 validation | PSNR: 26.69 -> 27.11 | Helped |
| Use TIC instead of PTI with 250 iterations | Image reconstruction · MS-COCO 2017 validation | PSNR: 26.71 -> 27.11 | Helped |
| Use TIC instead of PTI with 500 iterations | Image reconstruction · MS-COCO 2017 validation | PSNR: 26.92 -> 27.11 | Helped |
| Use TIC instead of DDIM reconstruction | Image reconstruction · MS-COCO 2017 validation | Time (s per image): 5.56 -> 5.56 | No clear effect |
| Use TIC instead of NTI with 250 iterations | Image reconstruction · MS-COCO 2017 validation | Time (s per image): 97.86 -> 5.56 | Helped |
| Use TIC instead of PTI with 250 iterations | Image reconstruction · MS-COCO 2017 validation | Time (s per image): 89.89 -> 5.56 | Helped |
| Add mask-guided attention concatenation to TIC | TIC general text-guided editing · Text-guided image editing | Content preservation and editability | Helped |
| Add ControlNet canny or depth control to mask-guided TIC | TIC general text-guided editing · Text-guided image editing | Layout and shape fidelity | Helped |
Limitations
- Mask-free TIC only reproduces input content and cannot generate new content absent from the input
- Shape of edited objects can drift slightly under mask-guided editing without ControlNet control
Experimental setup
Datasets: MS-COCO 2017 validation set
Benchmarks: MS-COCO 2017 validation (200 images)
Metrics: PSNR SSIM reconstruction time
Baselines: VAE reconstruction DDIM SDEdit Prompt-to-Prompt Null-Text Inversion Prompt-Tuning Inversion MasaCtrl
Builds on: DDIM Prompt-to-Prompt Null-Text Inversion MasaCtrl ControlNet
Closest papers by meaning
- Inversion-Free Image Editing with Language-Guided Diffusion Models (CVPR 2024 · similarity 0.93)
- DCI: Dual-Conditional Inversion for Boosting Diffusion-Based Image Editing (NeurIPS 2025 · similarity 0.91)
- Object-Aware Inversion and Reassembly for Image Editing (ICLR 2024 · similarity 0.90)
- Exploring Fixed Point in Image Editing: Theoretical Support and Convergence Optimization (NeurIPS 2024 · similarity 0.89)
- PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code (ICLR 2024 · similarity 0.88)
- EditInfinity: Image Editing with Binary-Quantized Generative Models (NeurIPS 2025 · similarity 0.88)
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.