Aeion Bridge + Stable Diffusion / Flux / ComfyUI

Aeion Bridge exposes local image generation as four MCP tools — generate (txt2img / img2img), edit (inpainting), upscale, and list-models — so cloud Aeion OS (and AeionClaw) can drive whatever Stable Diffusion, Flux, or ComfyUI-backed setup you run on your own GPU. Model support is whatever you have installed and running locally; Bridge doesn't dictate or bundle a specific model stack.

Supported Models

Bridge's generate/edit/upscale tools take a model string and pass it straight through — there's no fixed model allowlist on the platform side. Whatever checkpoint your local Stable Diffusion / Flux / ComfyUI setup has loaded is what Bridge will use. For reference, here's the landscape of what tenants commonly run behind Bridge:

Generating an Image via Bridge MCP

```typescript

Custom LoRAs — Bring Your Own, No Training Pipeline

Aeion doesn't run a LoRA training pipeline — there's no Bridge-side job that takes a reference-image dataset and produces a trained .safetensors file. If you train LoRAs (via Kohya_ss, or any other trainer, on your own machine or a rented GPU), the result plugs in at the generation step exactly like any other model: pass the LoRA's name/trigger in the prompt string, in whatever syntax your local backend understands (e.g. A1111-style <lora:name:weight> tags):

GPU Pool Management

```typescript

FAQ

Flux Schnell (Apache 2.0, fast, high-quality). Then add SDXL or SD 3.5 for specific use cases.

Flux Schnell (Apache 2.0), SDXL (CreativeML Open), SD 1.5 (CreativeML Open). Avoid Flux Dev (non-commercial) without Pro API.

Not today — there's no node-graph execution tool. Bridge exposes generate / edit / upscale / list-models; if you need a specific ComfyUI graph, you'd run it against your own ComfyUI server and use Bridge for the simpler generate/edit/upscale calls around it.

No — there's no training pipeline on the platform. Train on your own hardware (or a rented GPU) with your trainer of choice, then reference the result by name/trigger in the `prompt` if your local generation server supports that syntax.

Whatever your local inference server does with multiple GPUs is up to that server — Bridge doesn't do its own GPU scheduling or load-balancing across devices.

~$0.028/image effective (electricity + GPU amortization, RTX 4090, 3-year). Variable per-hardware and per-electricity-cost.

Keeping your local server's model already loaded (rather than cold-starting per request) is the biggest lever — that's a property of your generation server, not something Bridge configures.

Not a first-class Bridge parameter — if your local server supports ControlNet guidance, you'd configure that on the server side; Bridge's `generate` call doesn't have dedicated ControlNet fields.

Yes — `batchSize` on the generate call requests multiple images in one call (default 1).

Yes, in the sense that Bridge passes through any `model` string — but installing/managing checkpoint files on your machine is up to your local generation server, not something Aeion stores or manages.

Replicate Flux Schnell: $0.003/image. Aeion Bridge effective: $0.028/image (3-year amortization). Crossover at ~50K images/year (Bridge wins above).

Each tenant's Bridge instance is a separate connection to a separate machine — there's no shared model/checkpoint storage on the Aeion side to isolate.

Local image gen at production scale.