dragdropremove.com🪄

Custom Models

Have your own ONNX? Paste a Hugging Face link (both /blob and /resolve work). Supports .onnx and .onnx.zip files. The model downloads to this browser and is stored in IndexedDB for reuse. Your images are not uploaded.

Saved Recipes

Save your favorite combinations of model and settings for instant reuse. Each recipe gets an auto-generated name and color. Click the checkbox to load a recipe; tweak any setting to create a new variation. Share recipes with the copy button.

Choose what to display behind your final cutout. This fills the transparent areas in both images and videos.

CPU Threads (WASM only)

Choose how many CPU cores to use when WebGPU isn't available. More threads speed things up on multicore machines but may contend with other apps. Click "Apply" to recreate the session with your choice. (Ignored when WebGPU is active.)

Runtime Status

See which engine is running now. WebGPU uses your graphics card for peak speed; WASM runs on your CPU as a universal fallback. "Threads" shows the current CPU worker count.

Input Size (MODNet only)

MODNet is a trimap-free portrait matting model: instead of needing a user-drawn trimap, it predicts an alpha matte (0–1 transparency per pixel) in one forward pass. Internally, it learns coarse semantics (what looks like a person) and fine details (hair, edges), then fuses them into a clean matte. We run MODNet on a resized copy of your image, and then scale the resulting matte back up to your original resolution—your export size never changes.

What “Input Size” actually controls. This sets the shortest edge (in pixels) that we resize your image to before sending it to MODNet. A larger value gives the model more pixels to reason about tiny structures (individual hairs, semi-transparent strands), which usually yields a sharper matte—but uses more memory and runs slower. A smaller value is faster and lighter, but can miss very fine detail or produce stair-stepped edges.

Foreground Threshold

Decide how confident the model must be to treat a pixel as subject. Lower values include borderline hair and semi-transparent areas; higher values produce a tighter, safer cut.

Hard Cutoff

Convert the soft mask into pure black/white—no half-transparent pixels. Great for logos and products on flat backgrounds. When enabled, this overrides Edge Softness.

Edge Softness

Feather the mask edge to blend more naturally. Higher values ease transitions (nice for portraits); set to 0 for crisp, product-style edges.

Edge Contrast

Bias semi-transparent edge pixels. Values > 1.0 harden edges (more opaque); values < 1.0 soften them (more transparent). Use this when the cut looks too harsh or too wispy.

Sigmoid Activation

Only for custom models that output raw logits. Applies a sigmoid to convert logits into probabilities. Most bundled models already include this—leave off unless your model requires it.

Smoothing Radius

Set how far the guided filter looks when smoothing the mask while respecting image edges. Larger radii remove bumps; smaller radii keep fine detail.

Smoothing Sensitivity

Control how strictly the filter follows edges. Lower values cling to real edges (preserve detail); higher values allow smoothing to cross edges (cleaner, but risk blur).

Smoothing Strength

Blend between the original mask and the filtered one. 0 keeps the raw result; 1 fully applies smoothing. Use this to balance accuracy vs. cleanliness.

Shrink Mask (Erode)

Trim the mask inward to remove halos and tiny spikes. Each step removes about one pixel. Start small—too much erode can cut into your subject.

Grow Mask (Dilate)

Expand the mask outward to recover thin strands and fill tiny gaps. Each step adds about one pixel. Use sparingly to avoid pulling background back in.

Apply Shrink Before Grow

Choose operation order. Shrink → Grow removes noise first then restores size (cleanest). Grow → Shrink fills holes first then tightens edges (good for delicate structures).

Edge-Only Refinement Zone

Limit heavy refinement to a band around the mask edge. 0 processes the whole mask; 24–48 px targets the boundary for faster, cleaner results.

Color Spill Removal

Reduce green/blue fringing from backgrounds that bleed into hair and semi-transparent edges. Raise until halos fade; keep low for normal photos.

Sharpen Subject

Enhance edge definition in the cutout. 0.5–1.0 adds modest crispness; 1.5–2.0 is punchy but may show artifacts. Set to 0 if your photo is already sharp.

Result Smoothing Over Time

Blend results across frames for video or rapid tweaks. Higher values stabilize flicker but respond slower to change. Leave at 0 for still images.