Research map / Image generation / Editing and style transfer
Add-it: Training-Free Object Insertion in Images With Pretrained Diffusion 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.
Add-it inserts objects into real or generated images via a training-free extended multi-modal attention in FLUX, with weighted source/prompt/target attention, structure transfer, and subject-guided latent blending.
Problem
Adding objects to images from text instructions requires preserving the source scene while placing the new object at a plausible location (affordance), which existing diffusion-based editing and trained instruction-editing models handle poorly.
Approach
Extend MM-DiT self-attention in FLUX with source-image keys and values, rebalance attention among source, prompt, and target tokens via a key weight chosen by a root solver, add structure transfer via high-noise source latents, and blend latents using an SAM-2-refined attention mask.
Key results
- On the Additing Affordance Benchmark, Add-it reaches 0.828 affordance versus 0.474 for the best baseline (Prompt-to-Prompt).
- On Emu Edit add subset, Add-it obtains CLIP_dir 0.101, CLIP_out 0.322 and 81% inclusion, the best among compared methods.
- Human raters preferred Add-it in over 80% of head-to-head comparisons against baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use weighted extended attention with balanced source, prompt, and target contributions | Add-it · Additing Affordance Benchmark | Affordance: 0.474 -> 0.828 | Helped |
| Use weighted extended attention with balanced source, prompt, and target contributions | Add-it · Additing Benchmark | CLIP_dir: 0.170 -> 0.200 | Helped |
| Use weighted extended attention with balanced source, prompt, and target contributions | Add-it · Additing Benchmark | CLIP_out: 0.280 -> 0.261 | Hurt |
| Use weighted extended attention with balanced source, prompt, and target contributions | Add-it · Additing Benchmark | Inclusion: 97% -> 93% | Hurt |
| Use weighted extended attention with balanced source, prompt, and target contributions | Add-it · Additing Benchmark | CLIP_im: 0.958 -> 0.968 | Helped |
| Use weighted extended attention with balanced source, prompt, and target contributions | Add-it · Emu Edit | CLIP_dir: 0.091 -> 0.101 | Helped |
| Use weighted extended attention with balanced source, prompt, and target contributions | Add-it · Emu Edit | CLIP_im: 0.941 -> 0.929 | Hurt |
| Increase attention weight scale beyond the balanced automatic scale (e.g. gamma=1.2) (ablation) | Add-it · Additing Affordance Benchmark | Affordance | Hurt |
| Use the balanced automatic attention weight scale instead of the standard extended-attention scale (gamma=1.0) (ablation) | Add-it · Additing Affordance Benchmark | Affordance | Helped |
| Add Subject-Guided Latent Blending (ablation) | Add-it · Additing Benchmark | fine-detail preservation | Helped |
| Apply structure transfer too early in denoising (ablation) | Add-it · Additing Benchmark | Affordance | Hurt |
| Apply structure transfer later in denoising rather than at t=933 (ablation) | Add-it · Additing Benchmark | Inclusion | Hurt |
Surprising findings
- Simply appending source-image keys and values to extended attention causes the output to copy the source and ignore the edit prompt, requiring rebalancing.
- Applying the extended-attention mechanism to a real image with DDIM inversion fails on FLUX, motivating a noisy-source-latent alternative.
Limitations
- Relies on pretrained model biases, which may affect placement in unfamiliar or complex scenes.
- Requires detailed target prompts; a prompt like 'a dog' does not add another dog to a scene that already has one.
- Performance on real images is weaker than on generated images, attributed to the FLUX inversion method.
Experimental setup
Datasets: Emu Edit benchmark
Benchmarks: Emu Edit Additing Benchmark Additing Affordance Benchmark
Metrics: CLIP_dir CLIP_out CLIP_im Inclusion Affordance human preference
Baselines: InstructPix2Pix MagicBrush EraseDraw SDEdit Prompt-to-Prompt
Builds on: FLUX.1-dev Prompt-to-Prompt SDEdit SAM-2
Closest papers by meaning
- Paint by Inpaint: Learning to Add Image Objects by Removing Them First (CVPR 2025 · similarity 0.88)
- An Item Is Worth a Prompt: Versatile Image Editing with Disentangled Control (AAAI 2025 · similarity 0.88)
- Towards Efficient Diffusion-Based Image Editing with Instant Attention Masks (AAAI 2024 · similarity 0.87)
- RelationAdapter: Learning and Transferring Visual Relation with Diffusion Transformers (NeurIPS 2025 · similarity 0.87)
- Towards Understanding Cross and Self-Attention in Stable Diffusion for Text-Guided Image Editing (CVPR 2024 · similarity 0.87)
- Instruct-CLIP: Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning (CVPR 2025 · similarity 0.86)
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.