mlx-community/DDColor-paper-tiny-fp16

DDColor automatic image colorization (ConvNeXt-T backbone, ddcolor_paper_tiny.pth), converted to Apple MLX (-fp16) for Apple-Silicon inference via the mlx-ddcolor-swift Swift package.

Fastest tier (~0.1 s). Lightest backbone — good for batch / on-device.

A from-scratch MLX-Swift architecture port (no MLX donor) of DDColor: ConvNeXt encoder → quasi-UNet decoder → DETR-style multi-scale color decoder (100 color queries) → predicts the ab chroma channels, recombined with the source L (luminance). Grayscale / desaturated image in → colorized image at the same resolution out.

Use with mlx-ddcolor-swift (Swift / MLX)

// Package.swift → .package(url: "https://github.com/xocialize/mlx-ddcolor-swift", from: "0.1.0")
import DDColor

let colorizer = try DDColorColorizer.fromPretrained(
    weightsPath,                                   // model.safetensors from this repo
    config: DDColorConfig(tier: .large),           // .tiny for paper-tiny
    dtype: .float16)
let colorized: CGImage = colorizer(sourceCGImage)  // grayscale → color, source resolution

Or as an MLXEngine imageColorize ModelPackage (MLXDDColor.DDColorPackage), which resolves this repo automatically via the Hub.

Parity

Architecture parity-locked against the PyTorch oracle (piddnad/DDColor) on the CPU stream, fp32: max_abs ≤ 1.5e-6 (large) / 3.3e-6 (tiny) on the model's ab output. The cv2-exact LAB pre/post pipeline matches the reference ColorizationPipeline to ~9e-4 mean BGR error. This -fp16 build is visually identical to fp32 (cosine 1.000000, mean 0.04/255 on colorized output).

Weights: Apache-2.0 (piddnad/DDColor). Port code: MIT.