gaea-plugin-forge

The processing pack: ForgeCore's image models as one Gaea plugin, in one process. A worker costs about 146 MB, so small models share one. Each model is registered in the configuration ForgeCore's ForgeEngineService registers it with, under the id mlx-image-runtime's catalog gives it, so Gaea's image normalizer composes them with the Image profile's rules (exact ×scale, black-mask pixels kept, alpha re-applied, sizes) and chooses among them by its measured tiers. Add it by its repo URL:

gaea plugins add https://github.com/xocialize-code/gaea-plugin-forge

What it offers

MES operation	Models (catalog id)	Gaea's default
image.fill, image.erase, image.extend	lama-inpaint: LaMa, and MI-GAN as its fast mode (draft)	LaMa
image.upscale	nerve-upscale (×2 and ×4), realplksr-upscale (×4), heart-upscale (×4, slower), anime4k-upscale (Anime4K VL, instant), vosr2-upscale and nacre-upscale (generative, behind the text guard)	RealPLKSR; HEART at best. Anime4K, VOSR2 and Nacre by name
image.matte	birefnet-matte	BiRefNet
image.segment	edgetam-segment (fast), sam21-segment (SAM 2.1-Small)	SAM 2.1; EdgeTAM at draft
image.restore	nafnet-restore, restormer-motion, restormer-defocus, restormer-denoise, drunet-denoise, scunet-denoise, gfpgan-face, restoreformer-face	by kind: NAFNet for general, SCUNet for denoise, Restormer for deblur, GFPGAN for faces. DRUNet by name
image.relight	cidnet-relight (HVI-CIDNet)	CIDNet
image.colorize	ddcolor-colorize	DDColor
image.score	siglip2-quality (SigLIP2 no-reference IQA, 32 patches)	SigLIP2
helper x-gaea.image.defects	bopbtl-defects (BOPBTL scratch detection)	restore's scratches: the defects, then LaMa inpaints them

The worker maps one model call onto its engine capability (imageInpaint, imageUpscale, matting, promptSegment, imageRestore, imageRelight, imageColorize, imageQualityScore, defectDetect); Gaea's normalizer plans the job around it. Every model has one lane, in the quant ForgeCore runs it at. NERVE's, RealPLKSR's and NAFNet's weights ship inside the plugin, and Anime4K VL has none (Metal shaders); the others come from the models root Gaea sends in configure (MLXEngine's store by default), and the engine downloads a missing one on prepare.

VOSR2 and Nacre run behind ForgeCore's text-legibility guard, never bare, as pixel-agency-image-engine runs them (EngineImageEnhancer, AB-D-0111): the guard reads the input with the policy each pair earned (EngineTextGuardedUpscaler.measuredPolicy), then runs the generative tier, NERVE, or both composited under the words it protects (Nacre's policy protects words). The result names what made it (upscale_model). The worker links ForgeCore for this and for Anime4K VL's package, and registers only the pack's models.

Build and check locally

scripts/build-plugin.sh

scripts/check-plugin.sh

build-plugin.sh builds the worker in a scratch folder of its own and copies every SwiftPM resource bundle it links into the plugin's Contents/Resources, where the packages' bundled weights are found. check-plugin.sh restores the fixture photo with NAFNet over GPP, with every .build folder hidden; a control then shows that the bare worker, outside the bundle, can't find MLX's metallib.

swift build -c release --product forge-inproc && scripts/golden.sh <out-dir>

golden.sh calls every model over GPP on the signed bundle and compares its output with ForgeCore's own ForgeEngineService running the same call in-process (forge-inproc, on this package's one dependency resolution): 25 calls: 23 images byte-identical, and 2 scores the same value. The upscalers behind the guard are compared with ForgeCore's own EngineImageEnhancer (sign-256.png takes Nacre through word protection). scripts/make-fixtures.swift rebuilds the inputs 0.2.0 added.

scripts/release.sh builds, checks, notarizes with the fleet's mlx-notary profile, and publishes the GitHub release. See the demo plugin for signing, notarization and CI secrets.

Licence

The plugin is Apache-2.0 (Gaea G-21). Each model offer states its weights' licence from its package's manifest (Apache-2.0 or MIT for every model here); the ports are MIT or Apache-2.0. VOSR2's weights are Apache-2.0 on an undisclosed training corpus, which the fleet ships under posture A (AB-D-0105); Nacre is xocialize's, trained on Wikimedia Commons CC0, PD and CC BY images.