Connect your local AI

What it changes

If a Stable Diffusion already runs on your computer, Imaginode can use it like any catalog model: your checkpoints appear at the top of the model picker in Image nodes and every generation costs 0 credits. Your browser talks directly to your machine, nothing goes through our servers during generation.

Concretely: the final prompt (your text, completed by connected texts, @references and style) is sent to your local server, which generates the image. Imaginode then uploads it into your project's private storage, like an imported media file: it keeps its details (prompt, format, cost 0), its history, and can feed the other nodes, for instance a cloud Video node to animate it.

Supported engines

Two families are recognized and detected automatically on connection: ComfyUI, through its native API (workflow submission then job tracking), and AUTOMATIC1111's “/sdapi/v1” API, which Forge and SD.Next expose identically.

SwarmUI does not expose that API: if you use it, run a ComfyUI instead. Other applications are untested; those that faithfully mimic AUTOMATIC1111's API may work.

Setup: ComfyUI

Launch ComfyUI with the CORS permission:

python main.py --enable-cors-header https://imaginode.ai

Windows portable build: open run_nvidia_gpu.bat and add those two arguments at the end of the command line. The address to enter in Imaginode is http://127.0.0.1:8188.

The models offered are the files in your ComfyUI/models/checkpoints folder, the ones the “Load Checkpoint” loader sees. Split models loaded through UNETLoader (Flux as separate files, for instance) do not appear. Imaginode submits its own standard workflow: your saved workflows are neither read nor modified, and every generated image also stays in your ComfyUI/output folder.

Setup: AUTOMATIC1111, Forge, SD.Next

Linux and Mac:

./webui.sh --api --cors-allow-origins=https://imaginode.ai

Windows: open webui-user.bat and replace the line set COMMANDLINE_ARGS= with:

set COMMANDLINE_ARGS=--api --cors-allow-origins=https://imaginode.ai

Then restart. The address to enter is http://127.0.0.1:7860.

At generation time, Imaginode loads the chosen checkpoint through override_settings without restoring it afterwards: no model reload on every image when you chain generations.

Applied settings

Imaginode uses sober, universal settings: 28 steps, CFG 6, Euler a sampler (Euler ancestral on ComfyUI), random seed on every click. The Image node formats are translated into sizes aligned with the SDXL training buckets: 1:1 as 1024×1024, 16:9 as 1344×768, 9:16 as 768×1344, 4:3 as 1152×896, 3:4 as 896×1152.

If an image is wired to the node (start port, @reference or moodboard), the generation switches to img2img with a 0.6 strength. Only the first connected image is used as the start: there is no omni-reference locally, extra images are ignored.

Connection, privacy, credits

In a project, unfold the model picker of an Image node and click “Connect my local AI”: enter the address, test, your checkpoints show up along with the detected engine. The connection is remembered in this machine's browser, not in your account: on another device your local models do not appear, and a project referencing them will show a clear error at generation time instead of silently falling back to a paid model.

No credits are ever charged. During generation only your machine works; at the end, the image is uploaded into your project's private storage, exactly like a media file you import by hand. Delete it from the node to remove it from the project.

Troubleshooting

“Local AI unreachable”: three possible causes. The server is off. The CORS option was not passed at launch: the browser console (F12) then shows an explicit CORS error. Or you are not on the machine where it runs: a 127.0.0.1 address can only be reached from the computer itself.

On the first call, Chrome asks for permission to access devices on your local network: it must be accepted (it can be re-enabled in the site settings, icon left of the address bar). Safari blocks this kind of connection: use Chrome, Edge or Firefox. Each generation has 10 minutes to complete; beyond that, Imaginode shows an error, but your local server quietly finishes its work on its side.