Project
A local image generator and editor that runs on my own computer.
Where it runs: On my own computer
Online image tools send the prompt and the picture to someone else's server, and each one expects a prompt written the way its model likes it. I have a graphics card that can run current image models, and I wanted to know how far a simple, private tool could go on it: type an idea, switch a few things, get a picture, and edit it afterwards.
Imagine Image is a local picture maker and editor that runs on my own computer. I choose a local image model on a card, switch style chips and LoRAs on or off, pick a size and a quality, and type an idea. ComfyUI makes the picture on my graphics card. The Edit tab changes a picture: an area, an object, the background, the angle or the light. One button upscales the result. The whole app works offline.
Start the app with one script and open its page in the browser on the same computer.
On the Create tab, pick a model card, switch on LoRAs and style chips, choose an aspect ratio and Draft or Final quality, and type an idea.
Optionally let the prompt helper write the full prompt for the model I picked. It runs on a second computer on my home network.
Press Generate. ComfyUI makes one to four pictures, which land in the gallery strip.
Open a picture in the Edit tab to change it, or upscale it. Each edit is a new picture linked to the old one, with a before and after slider and an Undo.
One window with no page scroll. The model, up to six LoRAs, style chips, size, quality and seed are all cards, chips, sliders and buttons. Only the LoRAs that fit the chosen model are shown. Variations makes up to four pictures from one idea.
Prompt only is the default: my text goes to the model as I typed it. Engineer mode turns a short idea into a full prompt written for the active model, and Idea finder returns four idea cards to start from. Both use a language model on a second computer on my home network, and I can read and edit the prompt before I press Generate.
Change an area, remove an object, replace the background, expand a picture, restyle, relight, view from a new angle, fix faces and hands, and restore an old photo. Each edit is saved as a new picture linked to the old one.
Next to each picture the app saves a small recipe file with the prompt, model, seed, LoRAs and timings. Double-clicking a picture in the strip restores all its settings, so a result can be repeated or varied. Deleted items go to a trash folder first.
Speeds are measured on my own graphics card, not quoted from model pages: a Draft picture takes from under two seconds to about seven, depending on the model. The app has a Limits page that lists what does not work well yet, such as weak expansion on all four sides or small changes of angle.
A Python 3.11 app with FastAPI serves one static web page. A headless ComfyUI runs the picture workflows, started and stopped by the app. Before each job the app measures graphics memory and refuses a job that would not fit, with one plain line saying why. A shared lock makes sure only one of my local AI apps uses the graphics card at a time, so Imagine Image and AI Architect never collide. The prompt helper talks to Ollama on the second computer.
Hardware: an RTX 5090 workstation, with a Mac mini M4 as the prompt helper.
I built it by directing AI agents under written rules, with decisions logged and results measured on the real hardware. I set the requirements and reviewed the result.
Limits: it is for my personal use. It needs a Windows computer with an NVIDIA graphics card. The prompt helper needs the second computer to be on.
Stack
Still under construction
The Tsakmak brand and this website are still being built, so some pages and details are not final yet. Thank you for your patience.