Background Removal Showdown: RMBG-2.0 vs SAM 2 vs Proprietary APIs

Background removal is one of the most common image processing tasks in production. E-commerce platforms, headshot services, and creative tools all need it, and they need it to be fast, cheap, and accurate. We benchmarked the leading models and APIs to answer a simple question: is open source ready to replace paid APIs?

The short answer: yes. RMBG-2.0 scores within 2 points of the best proprietary API, at roughly 1/125th the cost per image.

Methodology

We curated a test set of 5,000 product images across three categories, each presenting different challenges for background removal:

• E-commerce products (2,000 images) — Clean studio shots, white/gray backgrounds, varied object shapes from fashion to electronics
• Headshots (1,500 images) — Professional portraits with complex hair edges, varied skin tones, and diverse backgrounds (offices, outdoors, gradients)
• Creative/complex (1,500 images) — Transparent objects, fine details (jewelry, lace), motion blur, low contrast foreground-background pairs

Every image was processed by all four models/APIs, and each output was scored on three metrics:

• IoU (Intersection over Union) — Measures mask accuracy against human-annotated ground truth
• Edge Quality — Custom metric evaluating the smoothness and precision of mask boundaries, weighted heavily around hair and fine details
• Transparency Handling — Accuracy on semi-transparent regions (glass, sheer fabric, smoke)

Models Tested

We evaluated four solutions, two open-source models self-hosted on our infrastructure, and two commercial APIs:

• RMBG-2.0 — BRIA AI’s latest open-source background removal model, running on A100 via Runflow
• SAM 2 — Meta’s Segment Anything Model 2, configured for automatic foreground segmentation, self-hosted
• Remove.bg API — Industry-standard commercial background removal API
• Proprietary C — A competing commercial API (anonymized per testing agreement)

Results: Composite Quality Score

The composite score weights IoU at 50%, edge quality at 30%, and transparency handling at 20%. All scores are on a 0–100 scale.

The gap between the best proprietary API (Remove.bg at 95) and the best open-source model (RMBG-2.0 at 93) is just 2 points. A year ago, that gap was closer to 8–10 points. Open source has closed the distance remarkably fast.

Results: Cost per Image

This is where the comparison gets dramatic. Self-hosting open-source models through Runflow costs a fraction of a cent per image, while API pricing runs between $0.02 and $0.05 per image at scale.

At 100,000 images per month, the cost difference is staggering: $40/mo with RMBG-2.0 on Runflow versus $5,000/mo with Remove.bg. That is a 125x cost difference for a 2-point quality gap.

Category Breakdown

The aggregate scores hide some interesting per-category dynamics:

• E-commerce products: RMBG-2.0 actually ties Remove.bg at 96. Clean backgrounds and well-defined objects play to RMBG’s strengths.
• Headshots: Remove.bg leads by 3 points (95 vs 92), primarily due to better hair edge handling on complex curly and fine hair.
• Creative/complex: Remove.bg leads by 4 points (93 vs 89). Transparent objects and low-contrast scenes remain the hardest challenge for open-source models.

Conclusion: Open Source Is Production-Ready

1. For e-commerce, switch now. RMBG-2.0 matches proprietary quality on product images at 1/125th the cost. There is no reason to keep paying API prices.
2. For headshots, it depends on your quality bar. The 3-point gap on hair edges matters for premium portrait services. For most use cases, RMBG-2.0 is more than sufficient.
3. For complex creative work, APIs still have an edge. But the gap is closing fast, and the cost savings from open source fund a lot of post-processing cleanup.
4. SAM 2 is versatile but not specialized. It is a general-purpose segmentation model, not purpose-built for background removal. RMBG-2.0 wins on this specific task.

We will re-run this benchmark quarterly as new model versions are released. All test images, scoring code, and raw results are available in our open benchmark repository.