# AI Background Removers Tested: Bria RMBG 2.0 vs Rembg (2026)

Canonical page: https://fuser.studio/articles/best-ai-background-removers

We ran Bria RMBG 2.0 and Rembg on backlit curly hair and clear glassware. See the edge differences, what neither handles, and when to generate a transparent image or replace the background instead.

[All guides](/articles) · [Best AI image editing models](/articles/best-ai-image-editing-models)

**Quick answer:** for most cutouts, use [Bria RMBG 2.0](/models/background-remover), which Fuser ships as the **Background Remover** node. In our test it separated backlit curly hair from a park background and kept every part of a glass perfume bottle. [Rembg](/models/rembg) costs fewer credits and is fine for a clear subject on a plain background, but it left a green halo in the hair and deleted the bottle's gold cap. Neither tool makes clear glass see-through. If you need a transparent product image and don't have a photo yet, generate one with [GPT Image](/models/gpt-image) and its Transparent toggle. If you only want a different backdrop, use the [Background Replacer](/models/background-replacer).

## What Fuser offers for background removal

Fuser has two dedicated removers. Both take one image and return a PNG with an alpha channel.

- **Background Remover** runs [Bria RMBG 2.0](https://fal.ai/models/fal-ai/bria/background/remove). Bria's [model card](https://huggingface.co/briaai/RMBG-2.0) says it is built on the BiRefNet architecture and trained on more than 15,000 manually labelled, fully licensed images. It returns an 8-bit alpha matte, so edges can be partly transparent instead of simply on or off. The node has no settings.
- **Rembg** runs fal's [imageutils/rembg](https://fal.ai/models/fal-ai/imageutils/rembg) endpoint. Its one control, **Crop**, trims the output to a box around the subject. The open-source [rembg project](https://github.com/danielgatis/rembg) supports many segmentation models, and fal's page does not say which one this endpoint uses, so judge it by its output rather than by the project's feature list.

Two other nodes help with the same job. The **Background Replacer** ([Bria](https://fal.ai/models/fal-ai/bria/background/replace)) puts the subject in a new setting from a text prompt or a reference image. **GPT Image** can generate an image with a transparent background from text.

## Test: backlit hair and clear glass

Hair and glass are the two classic failure cases, so we generated two photos that contain both: a woman with backlit, frizzy curls holding a wine glass in a leafy park, and a tumbler, a clear perfume bottle and a potted succulent on a café table. Each photo went through both removers at default settings, once each. We placed every cutout on a flat navy layer, because leftover background is hard to see against a checkerboard.

![Two source photos and their cutouts from Bria RMBG 2.0 and Rembg on a navy background: Rembg leaves a green halo around the hair and drops the perfume cap.](https://statics.fuser.studio/cms/907b5426-7dc4-41fd-8b87-6ed249a77946)

_Same two 1024 px photos through both removers, default settings, one run each. Cutouts placed on navy so leftover background is visible. Run on 28 September 2026._

**Hair.** Bria RMBG 2.0 kept individual backlit strands and removed the park between them. About half its pixels came back fully opaque and the rest were fully or partly transparent, which is what a soft hair edge needs. Rembg traced the outline of the hair but kept a band of green foliage between the curls. On navy that shows as a halo you would have to paint out by hand.

**Glass.** Bria kept the tumbler, the lemon, the perfume bottle with its gold cap, and the plant, with clean edges on each. Rembg removed the gold cap completely (its alpha there is zero), and left a scrap of tabletop under the tumbler and a sliver beside the bottle.

![Close-up crops: Bria RMBG 2.0 cuts backlit hair strands cleanly; Rembg keeps foliage between curls, removes the perfume cap and leaves a table fragment under the tumbler.](https://statics.fuser.studio/cms/44579c4c-a40a-4e27-b417-363a14a47afd)

_Crops from the cutouts above._

**What neither did.** Both removers treated the glass as solid. The warm tabletop and café colours that showed through the glass in the original photo are still inside the cutout, fully opaque, so the bottle looks tinted when you put it on a new background. This is a limit of cutting out an existing photo. The light that passed through the glass is already baked into the pixels.

This is one test on two images, not a benchmark, but the failure modes match what each tool is built for. Bria RMBG 2.0 was trained on high-resolution, pixel-labelled images of people and objects. Rembg is a general-purpose open-source tool.

## When a remover is the wrong tool

**You don't have the photo yet.** Generate the product with transparency built in. In Fuser, turn on **Transparent** in the GPT Image node. The toggle only works for text-to-image. It is ignored once you connect an input image, so GPT Image can't cut out a photo you already have. That restriction comes from how Fuser's node builds the request. On the API side, OpenAI's [image generation guide](https://developers.openai.com/api/docs/guides/image-generation) says a transparent background needs the background parameter set to transparent and PNG or WebP output. Our GPT Image 2.5 Flare run returned a perfume bottle with about 65% of the canvas fully transparent. Like the removers, it rendered the glass body as opaque, not see-through.

**You want a new scene, not a cutout.** The Background Replacer generates a new background around the subject in one step. It kept the tumbler and bottle and put them on a stone plinth, but it treated the succulent as background and removed it. Check that every object you need is still in the frame. For more control over what changes, an instruction editor such as [Qwen Image Edit](/models/qwen-image-edit) or [FLUX.1 Kontext](/models/flux-1-kontext) lets you describe the new setting in words. See [the best AI image editing models](/articles/best-ai-image-editing-models). Editors return a flat image, not an alpha channel.

![A transparent-background perfume bottle generated by GPT Image, and the glassware photo on a new stone plinth from Bria Background Replacer, which also removed the plant.](https://statics.fuser.studio/cms/6118b142-61f4-41ff-8337-0b93c86deb32)

_Left: GPT Image 2.5 Flare with background set to transparent, the value Fuser's Transparent toggle sends. Right: Background Replacer with a text prompt. Both run on 28 September 2026._

**You need the cutout on a precise layout.** Send the Background Remover output into the [Compositor](https://docs.fuser.studio/docs/compositor/layers), put a colour fill or another image layer underneath, and export a PNG at the canvas size you set.

## Build it once in Fuser

Connect your product or portrait to a Background Remover node and route the transparent PNG wherever it goes next: the Compositor for layouts, an [image-to-video model](/articles/best-image-to-video-models) for an animated ad, or a [product photography workflow](/articles/best-ai-product-photography-tools). Branch the same image into Rembg as well if you want a cheaper pass to compare against. Save the graph as a [Recipe](/features/recipes) and every new product goes through the same steps. For a full example, see [product photo to video ad](/articles/product-photo-to-video-ad).

Last verified September 28, 2026.

## Pick the right node for the cutout.

What each option did in our test, and when to use it.

### Remove a background

| Node | Use it when | Watch out for |
| --- | --- | --- |
| Background Remover (Bria RMBG 2.0) | Hair, fur, products and commercial work; licensed training data. | Glass keeps the original background's colour inside it. |
| Rembg | Clear subjects on plain backgrounds, when credits matter. | Left a halo in hair and deleted a small gold cap in our test. |

### Alternatives

| Node | Use it when | Watch out for |
| --- | --- | --- |
| GPT Image (Transparent on) | You're generating the product from text and need a PNG with alpha. | Text-to-image only; can't cut out an existing photo. |
| Background Replacer (Bria) | You want the subject in a new scene in one step. | Can treat secondary objects as background and remove them. |
| Instruction editors | You want to describe the new setting in words. | Return a flat image with no alpha channel. |

## Questions, answered.

### What is the best AI background remover?

Of the two removers in Fuser, Bria RMBG 2.0 (the Background Remover node) gave the cleaner result in our test. It kept backlit hair strands without leftover background and kept every part of a glass perfume bottle. Rembg left a green halo in the hair and removed the bottle's cap.

### Can AI remove the background from hair?

Yes. Look for a tool that outputs a soft alpha matte rather than a hard mask. Bria RMBG 2.0 returns an 8-bit alpha matte, and on backlit curly hair it separated individual strands. Check the result on a dark, flat colour, because leftover background is hard to see on a checkerboard.

### Can AI background removers handle glass?

They can cut glass out cleanly, but in our test neither Bria RMBG 2.0 nor Rembg made it see-through. The table and room colours visible through the glass stay inside the cutout. For a transparent product shot, generate it with a transparent background or shoot it on the final backdrop.

### Can GPT Image remove a background from my photo?

No. GPT Image's transparent background works only when it generates an image from text. Once you connect an input image, the setting doesn't apply. Use the Background Remover node to cut out an existing photo.

### What file format do background removers output?

Both Fuser removers return a PNG with an alpha channel. In our test, a 1024 px input came back as a 1024 px PNG from each. Rembg can also crop the output to the subject's bounding box.

### Is Bria RMBG 2.0 safe for commercial use?

Bria says RMBG 2.0 was trained only on licensed images. The open weights are licensed CC BY-NC 4.0, and commercial use of the weights needs an agreement with Bria. In Fuser you run it as a hosted model through fal.

## Cut out, place and animate on one canvas.

Run both removers side by side, then send the best cutout to the next step.

[Try the Background Remover](https://fuser.studio/models/background-remover) · [Explore all guides](https://fuser.studio/articles)

## More articles

- [Best AI Logo and Vector Generators (2026): SVG vs PNG, Tested](https://fuser.studio/articles/best-ai-logo-and-vector-generators.md)
- [FLUX.1 Kontext Editing Guide: Instructions, Identity and Iterative Edits](https://fuser.studio/articles/flux-kontext-editing-guide.md)
- [Instagram Post Sizes (2026): Feed, Story, Reel and Carousel](https://fuser.studio/articles/instagram-post-sizes.md)
