remove_background/src/sessions/mod.rs

use image::{DynamicImage, GenericImageView, GrayImage, Luma, imageops::FilterType};
use ndarray::{Array4, IntoDimension};
use ort::{session::Session as OrtSession, value::Tensor};
use std::error::Error;

mod birefnet_lite;
mod bria;

pub use birefnet_lite::BiRefNetLiteSession;
pub use bria::BriaSession;

/// Common interface for background-removal models, mirroring rembg's
/// `BaseSession` pattern. Each concrete session owns an `ort::Session` and
/// implements `inner` / `input_name`; the rest (preprocessing, inference,
/// postprocessing into a mask) is provided by default implementations.
pub trait Session {
    /// Canonical model name (matches rembg's filename stem).
    fn name() -> &'static str
    where
        Self: Sized;

    /// URL to download the ONNX model from.
    fn url() -> &'static str
    where
        Self: Sized;

    /// Optional SHA-256 checksum used to verify the cached model.
    fn sha256() -> Option<&'static str>
    where
        Self: Sized,
    {
        None
    }

    fn mean(&self) -> (f32, f32, f32) {
        (0.485, 0.456, 0.406)
    }

    fn std(&self) -> (f32, f32, f32) {
        (0.229, 0.224, 0.225)
    }

    fn input_size(&self) -> (u32, u32) {
        (1024, 1024)
    }

    /// Whether a sigmoid should be applied to the raw logits before the
    /// min/max normalization step. `birefnet-*` needs this; `bria-rmbg` does not.
    fn apply_sigmoid(&self) -> bool {
        false
    }

    fn inner(&mut self) -> &mut OrtSession;

    fn input_name(&self) -> &str;

    /// Port of rembg's `BaseSession.normalize`: resize with LANCZOS,
    /// scale into `[0, 1]` by dividing by the max pixel value, then apply
    /// channel-wise mean/std.
    fn normalize(&self, img: &DynamicImage) -> Result<Array4<f32>, Box<dyn Error>> {
        let (w, h) = self.input_size();
        let resized = img.resize_exact(w, h, FilterType::Lanczos3).to_rgb8();
        let (width, height) = resized.dimensions();

        let mut max_pixel: f32 = 0.0;
        for p in resized.pixels() {
            for c in 0..3 {
                let v = p[c] as f32;
                if v > max_pixel {
                    max_pixel = v;
                }
            }
        }
        let denom = max_pixel.max(1e-6);

        let mean = self.mean();
        let std = self.std();

        let mut array = Array4::<f32>::zeros((1, 3, height as usize, width as usize));
        for y in 0..height {
            for x in 0..width {
                let pixel = resized.get_pixel(x, y);
                let r = pixel[0] as f32 / denom;
                let g = pixel[1] as f32 / denom;
                let b = pixel[2] as f32 / denom;
                array[[0, 0, y as usize, x as usize]] = (r - mean.0) / std.0;
                array[[0, 1, y as usize, x as usize]] = (g - mean.1) / std.1;
                array[[0, 2, y as usize, x as usize]] = (b - mean.2) / std.2;
            }
        }

        Ok(array)
    }

    /// Run inference and return a grayscale mask resized to the input image.
    /// Mirrors rembg's `predict`:
    ///   1. `inner_session.run(normalize(img, mean, std, size))`
    ///   2. take the first output, channel 0
    ///   3. optional sigmoid (birefnet)
    ///   4. min/max normalize into `[0, 1]`
    ///   5. scale to `u8`, resize to original image dimensions
    fn predict(&mut self, img: &DynamicImage) -> Result<GrayImage, Box<dyn Error>> {
        let (orig_w, orig_h) = img.dimensions();
        let input = self.normalize(img)?;
        let apply_sigmoid = self.apply_sigmoid();

        let input_name = self.input_name().to_string();
        let outputs = self
            .inner()
            .run(ort::inputs![input_name => Tensor::from_array(input)?])?;

        let output = &outputs[0];
        let (shape, data) = output.try_extract_tensor::<f32>()?;

        if shape.len() != 4 {
            return Err(format!(
                "Expected 4D output tensor [N, C, H, W], got shape {:?}",
                shape
            )
            .into());
        }
        let (n, _c, h, w) = (
            shape[0] as usize,
            shape[1] as usize,
            shape[2] as usize,
            shape[3] as usize,
        );
        if n != 1 {
            return Err(format!("Expected batch size 1, got {}", n).into());
        }

        let view = ndarray::ArrayView::from_shape(
            (
                shape[0] as usize,
                shape[1] as usize,
                shape[2] as usize,
                shape[3] as usize,
            )
                .into_dimension(),
            data,
        )?;

        // Take channel 0: pred = out[:, 0, :, :]
        let mut pred: Vec<f32> = Vec::with_capacity(h * w);
        for y in 0..h {
            for x in 0..w {
                let mut v = view[[0, 0, y, x]];
                if apply_sigmoid {
                    v = 1.0 / (1.0 + (-v).exp());
                }
                pred.push(v);
            }
        }

        let (mut mi, mut ma) = (f32::INFINITY, f32::NEG_INFINITY);
        for &v in &pred {
            if v < mi {
                mi = v;
            }
            if v > ma {
                ma = v;
            }
        }
        let range = (ma - mi).max(1e-6);

        let mut mask = GrayImage::new(w as u32, h as u32);
        for y in 0..h {
            for x in 0..w {
                let v = (pred[y * w + x] - mi) / range;
                let u = (v.clamp(0.0, 1.0) * 255.0).round() as u8;
                mask.put_pixel(x as u32, y as u32, Luma([u]));
            }
        }

        let mask = image::imageops::resize(&mask, orig_w, orig_h, FilterType::Lanczos3);
        Ok(mask)
    }
}