gaea-plugin-ming-image

inclusionAI's Ming-Image 0.1 Design (MIT) as a Gaea plugin: a Ling-mini MoE MLLM conditioning a Z-Image-shaped DiT, with an RGBA VAE. It is a design model, for posters, signage, layouts and text on the page, from natural language or a JSON layer schema. It is built on GaeaPluginKit over ming-image-swift, runs in its own process, and speaks the Gaea Plugin Protocol v1. Add it to Gaea by its repo URL:

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

What it offers

Model	MES operation	Quality tiers	Notes
ming-image	image.generate	standard, best	12 steps without CFG (the reference's Design profile). Not a default: Gaea's routes don't change, so a request reaches it by naming it.

• Sizes. Ming-Image draws in two buckets, 1024² and 2048²: a request takes the smallest bucket that covers its area, at that bucket's nearest aspect, resized to cover the requested aspect: a 768×512 request is drawn at 1248×832. The worker reports the size drawn in effective, with size_snapped; Gaea then returns the size the request asked for. Up to 2048² in area, at most 8:1.
• count: 1–8 images, each with seed + i. A request without a seed gets one, reported in effective.
• Options, as MES extensions: x-ming.steps (1–50), x-ming.guidance (1–10, 1 turns CFG off) and x-ming.background (below).
• Transparent backgrounds (0.2.0): x-ming.background: "transparent" draws the design on native alpha and returns a straight-alpha PNG, for stickers, logos and overlays. The package prepends its measured phrase, retries once at the same seed with a second phrase when the first image comes back opaque (about twice the time), and floors α below 0.1 to 0. The worker checks every image by the package's gate metric (more than 10% of pixels with α ≤ 12), and one that is still opaque carries the warning background_opaque. Scenes that fill the frame do; so, sometimes, does glass. image.matte can cut a subject out of such an image.
• Negative prompts do nothing (negative_prompt_ignored): the model conditions on the prompt alone.

Layers: ming-image-layer (0.3.0)

Model	MES operation	Quality tiers	Notes
ming-image-layer	image.decompose	draft, standard, best	Ming-Image 0.1 Design-Layer: a flat design into up to 12 straight-alpha RGBA layers, front-most first, plus its own composite. 12 steps, guidance 2 (the reference's Layer profile).

• The plan is the prompt, front-most first: "Decompose this image into 3 layers. Layer 1: the words … Layer 2: the saxophone … Layer 3: the background." Without one the model chooses, and its split is much rougher: on the jazz poster, 4 unplanned layers repeated the saxophone and left one nearly empty, where a 3-layer plan gave the headline, the saxophone and the background, one each.
• layer_count: 1–12 (limits.max_layer_count). A plan's stated count must agree with it. Without either, the package makes 4 layers, or 5 with a plan that states none; effective.layer_count says how many came back.
• Sizes. The package works at the input's aspect in its 1024 bucket, or 512 at quality: draft (resolution in effective), and returns that size. Gaea's normalizer brings the layers to the input's size.
• outputs: layers, and composite on request (supports.composite).
• Speed: about 6 minutes for 4 layers at 1024 (int8 or bf16) and about a minute at 512, on an M5 Max.

Lanes and memory

Lane	Resident	Peak while running	Gaea schedules
int8 (default)	7.9 GB	24.4 GB	32.3 GB
bf16	13.5 GB	44.1 GB	57.6 GB
int4	7.9 GB	14.6 GB	22.5 GB

Every lane peaks while the MLLM conditions the prompt: it loads for each job, conditions, and is freed before the denoise. int8 is the default, as in mlx-image-runtime's catalog; speed is within about 10% across lanes (the DiT is compute-bound, and 8-bit in every tier), so the lanes trade memory and fidelity. When int8 doesn't fit, Gaea loads a smaller lane and warns (lane_fallback, G-23); gaea models lane ming-image int4 makes another lane the default.

Weights

Gaea sends its models root in configure (MLXEngine's store by default). Each lane reads its published tier, mlx-community/Ming-Image-0.1-Design-{8bit,bf16,4bit}, there and downloads it on prepare or materialize when it's missing. The package refuses a snapshot whose quantization isn't the lane's, or the Design-Layer checkpoint.

Build and check locally

scripts/build-plugin.sh

scripts/check-plugin.sh

build-plugin.sh copies every SwiftPM resource bundle into the plugin, not only MLX's metallib: Ming-Image's tokenizer config names no tokenizer class, so swift-transformers reads its bundled fallback configs through Bundle.module, and without the bundle the worker traps. check-plugin.sh generates one 1024² image at int8 in 6 steps against the weights in MLXEngine's store, with every .build folder hidden, then shows that the bare worker outside the bundle finds no metallib. scripts/release.sh builds, checks, notarizes with the fleet's mlx-notary profile, and publishes the GitHub release.

ming-inproc (its header lists the options) is not part of the plugin: it runs the same lane in-process through MLXServeEngine, the reference Gaea's runs are compared against.

scripts/golden.sh <out-dir>

golden.sh runs three seeded jobs over GPP on the signed bundle and through ming-inproc and needs each pair the same: int8 at 1024², int8 at 1280×720, int8 at 768×512 (drawn at 1248×832), and int4 at 1024².

Licence

The plugin is Apache-2.0 (Gaea G-21). Ming-Image 0.1 Design's weights are MIT (inclusionAI), and ming-image-swift is MIT.