AIAny
AI Model2026
Icon for item

RMBG-3.0

Turns photos into detailed foreground alpha mattes, with particular attention to hair, fur, shadows, and complex edges. Built on a generative image-editing model and distilled to four steps; self-hosted weights are non-commercial only and require substantial GPU memory.

Introduction

Background removal is often treated as tracing an object's outline. That breaks down when the subject is semantically ambiguous—a cast shadow, a strand of hair, or a gap in a bicycle wheel. This model takes a different route: it treats the task as an image edit and renders an alpha matte, aiming to decide what belongs to the subject rather than merely follow visible edges.

Key Capabilities
  • Its generative foundation is an 8B-parameter Fibo 2.0 model fine-tuned for alpha matting. The approach is intended to handle multiple subjects, motion blur, hair and fur, holes in mesh, and shadows more thoughtfully than conventional segmentation.
  • The model is distilled to four inference steps, without classifier-free guidance. Bria reports about six seconds per roughly 1-megapixel image on an L40S, with peak memory around 28.5 GiB; practical self-hosting therefore calls for a high-memory GPU.
  • Its output is a single-channel matte: white marks the foreground, black the background, and grey represents partial transparency. This preserves soft boundaries that a hard binary mask would discard.
  • In Bria's comparison on 234 deliberately difficult images, it ranked first overall among six tested systems. This is a vendor-reported benchmark, not an independent universal ranking, and results vary by image type.
Who It Suits—and the Trade-offs

A good fit for quality-first product imagery and difficult cutouts where edge detail matters more than throughput. Look elsewhere for high-volume, low-latency processing, small GPUs, or scenes demanding convincingly transparent glass, nets, and smoke: the model can render these too opaquely, occasionally includes nearby clutter, and has rare failed outputs. The Hugging Face weights use CC BY-NC 4.0; commercial self-hosting requires a separate agreement with Bria AI.

Information

Categories

More Items

Hugging Face
AI Model2026

Transcribes speech audio into text across 45 languages, with automatic language detection. Its 0.5B parameters make it a compact option for multilingual transcription using Transformers or vLLM.

Hugging Face
AI Model2026

Combines Korean-English language and image understanding with screenshot-driven browser control in one model. It prioritizes retaining general LLM and VLM capabilities rather than optimizing solely for web tasks.

Hugging Face
AI Model2025

Answers questions and generates descriptions from video, images, and text. Video inference uses about 1.38 GB of GPU memory, making it suitable for resource-constrained devices and task-specific fine-tuning.