Image Edit - LatentDiffusion(Object Removal)

What it does

Image Eraser removes unwanted objects from your photos and intelligently fills the space with realistic background. Instead of cropping or leaving a blank spot, it uses AI to understand the surrounding context and generate a natural-looking replacement. You simply mark what you want removed (using a mask), and the model regenerates that area as if it was never there. Works on any type of photo—people, objects, nature, architecture, anything.

Problem it solves

• Photobomb removal – That random person in the background of your photo
• Unwanted objects – Power lines, poles, litter, signs, vehicles ruining the shot
• People removal – Ex-partner, stranger, or random person in group photos
• Blemishes & defects – Scratches, spots, or marks on photos and documents
• Shadow/reflection cleanup – Unwanted shadows, reflections, or light artifacts
• Social media & publishing – Clean up photos for posts, articles, product images
• E-commerce & real estate – Remove clutter from product/property photos
• Professional editing – Efficient alternative to manual photoshop work
• No special tools needed – Just draw a rough outline of what to remove (mask)

Input/Output

• Input: Original photo + mask

1. Photo: Any RGB photo (JPEG, PNG, etc.)

• Mask: A rough outline of what to remove (same size as photo)
• Tools: Image Masker / Image Masker by Class / Image Masker by Coord
• White/bright = regions to erase (the "bad" parts)
• Black/dark = regions to keep (the "good" parts)
• Can be rough—the AI figures out the exact details
• An example to create a Mask (with Image Masker in this case)
• Add an Input Image and connect with Image Masker

• Create Mask
• Brush regions to erase and Apply Mask when done

• Run Flow to download a Mask image

• Random Seed: Any number; controls the variation in generated content (same seed = same result)
• The number of diffusion steps: 10–100 (more steps = higher quality but slower)
• 10-20 steps: Fast preview
• 50 steps: Good balance (quality & speed)
• 100 steps: Slower but best quality
• Output: The same photo with object removed
• Same resolution as input
• Unwanted region replaced with realistic, context-aware background
• Seamlessly blended with surrounding content
• RGB image (JPEG, PNG, or tensor)

Accuracy & Speed

• Quality Metrics (from paper benchmarks):

• Quality Characteristics:
• Small objects (photobombs, small people): Excellent removal, realistic fill
• Large areas: Good results, may show subtle blending artifacts
• Complex textures (grass, water, patterns): Usually very good, occasionally needs refinement
• Faces & people: Good results, may need touch-ups in some cases
• Shadows/reflections: Good removal, realistic lighting
• Hard edges (architecture, lines): Can sometimes be visible, may need post-processing

Technical Details

• Architecture: Latent Diffusion Model with Inpainting
• Core Components:
• VAE Encoder: Compresses images to latent space (16-64× faster than pixel-space diffusion)
• Diffusion Process: Denoises latent representations using mask + image context
• Inpainting Conditioning: Mask and surrounding context guide the diffusion process—no retraining needed for different mask shapes
• Sampling:
• Diffusion steps: 10-100 (more steps = higher quality)
• DDIM sampler with configurable stochasticity (eta: 0.0 = deterministic, 1.0 = varied)
• Classifier-free guidance: Strength 5.0-7.5 (higher = more faithful to mask)
• Random seed: For reproducible results
• Requirements:
• PyTorch (v1.8+)
• NVIDIA GPU with 8GB+ VRAM
• Training:
• Pre-trained VAE and diffusion model
• Trained on masked image regions with surrounding context

Technical Details

🛡️ Compliance & Provenance

For more on how we classify models and mark outputs, see our AI Policy.

Model Source

• GitHub Repository: https://github.com/CompVis/latent-diffusion (Official implementation with pretrained checkpoints and inpainting scripts)
• Paper: Rombach et al., "High-Resolution Image Synthesis with Latent Diffusion Models" (CVPR 2022, arXiv:2112.10752) https://arxiv.org/abs/2112.10752
• License: MIT License