GAI-A user manual

Working with GAI-A

A step-by-step guide to the GAI-A interface: the Create page, its four workflows, and the parameters that control each generation.

Open the interface

workflows Blend, Canny, Depth, Text
model Stable Diffusion via ComfyUI
requires a recent web browser

01Overview

GAI-A is a goal-oriented web interface that lets designers work with Generative AI models from their browser, with no installation or setup. Behind the interface, each workflow runs a ComfyUI pipeline built on Stable Diffusion. You only see the inputs and settings that matter for the task.

The interface currently offers four workflows, each tailored to a specific use:

Blend img2img: Blends a collage of input images into a single image, with a varying degree of abstraction. Use it to turn references and fragments into one design idea.
Canny controlnet: Detects the edges of an input image and generates a new image that follows those edges. Use it for sketches, line drawings and facades.
Depth controlnet: Estimates the depth of an input image and generates a new image that keeps its spatial structure. Use it for model views and massing studies.
Text txt2img: Generates an image purely from text prompts. Use it to explore an idea before you have any images.

02Getting started

2.1Open the homepage

Go to interface.gaia.itu.edu.tr. The homepage introduces the interface and links to the About page, where you can follow its development.

The GAI-A homepage with the title “Generative AI in Architecture” and a Create button — **Fig. 1**The GAI-A homepage.

2.2Go to the Create page

Click Create in the menu or below the title, or go directly to interface.gaia.itu.edu.tr/create. All generation happens on this page, through a canvas application.

The homepage with the Create link in the menu and the Create button highlighted — **Fig. 2**Either Create button opens the Create page.

03The Create page

The Create page is laid out around a central workspace. Its layout stays the same in every workflow, but some fields and parameters appear or disappear depending on the workflow you select. Read this section and look over the page before you start generating.

The empty Create page: prompt boxes on the left, the workspace frame in the middle, instructions on the right — **Fig. 3**The Create page with the Blend workflow selected.

3.1Prompt boxes

The two boxes on the left of the page hold your prompts. The green Positive box describes what you want to see; the red Negative box lists what the model should avoid.

The positive and negative prompt boxes highlighted on the left of the page — **Fig. 4**The prompt boxes.

Keep prompts short and concrete: materials, setting, and the kind of image you want, such as a sketch, a render or a photograph. In the example below, the positive prompt asks for a timber and concrete facade drawn as a sketch, and the negative prompt keeps people and trees out of the result.

A positive prompt reading “Concrete and wood facade, architectural sketch” and a negative prompt reading “People, trees” — **Fig. 5**A positive and a negative prompt.

3.2Workflow toggles

The workflow toggles sit in the bottom right corner. The canvas changes slightly depending on the workflow you select.

The workflow toggles Blend, Canny, Depth and Text highlighted in the bottom right corner — **Fig. 6**The workflow toggles.

3.3The workspace

The grey frame in the middle of the page is the workspace. Only what lies inside the frame is sent to the model.

The grey workspace frame and the Upload Image and Generate buttons highlighted in the middle of the page — **Fig. 7**The workspace frame, with the Upload Image and Generate buttons below it.

Drag and drop images onto the canvas, or click Upload Image to choose a file from your computer. Select an image to resize, rotate, reorder or delete it; the instructions on the right of the page list the shortcuts. PNG files with transparent backgrounds keep their transparency.

A collage of a tree placed over an architectural sketch inside the workspace, with selection handles — **Fig. 8**A collage in progress: a tree placed over a sketch, with the tree selected.

important

Empty areas inside the frame are converted to black and affect the result. If your composition does not fill the frame, place a background image at the bottom of the stack.

Click Generate to send the workspace to the model. The result appears on the canvas when it is ready. Click Clear to remove it and return to your composition.

A generated image of a concrete and timber building shown on the canvas, with a Clear button below — **Fig. 9**A generated result on the canvas.

3.4Outputs

The Outputs button in the top left corner opens your generation history.

The Outputs button highlighted in the top left corner — **Fig. 10**The Outputs button.

The Outputs tab slides in from the left and lists every image generated in the session. Click an image to see it full screen.

The Outputs tab open on the left, listing four generated images — **Fig. 11**The Outputs tab.

3.5Advanced options

The Advanced Options button in the top right corner opens the parameter panel.

The Advanced Options button highlighted in the top right corner — **Fig. 12**The Advanced Options button.

The panel shows the parameters of the selected workflow, each with a help icon. The values below are the defaults for Blend. Each parameter is explained in Advanced parameters.

The advanced options panel with sliders for Steps, CFG, Denoise, Relative Output Height and Width, and an Iterate checkbox — **Fig. 13**The advanced options panel in the Blend workflow.

04Keyboard shortcuts

These actions work on the image selected in the workspace.

Action	Shortcut
Select or deselect an image	Click
Send the image back	`Ctrl` + `B`
Bring the image forward	`Ctrl` + `M`
Delete the image	`Del`
Upload an image	Drag and drop onto the canvas

05Basic examples

Work through these two examples to get to know the interface.

5.1Text to image

Select the Text workflow.
Enter a positive prompt, for example “Concrete and glass building near a lake, photograph”, and the negative prompt “People”.
Click Generate.
Open the Outputs tab to see your generation history.

A glass and concrete building by a lake generated with the Text workflow, with earlier results in the Outputs tab — **Fig. 14**Result of the text-to-image example, with earlier generations in the Outputs tab.

5.2Image to image

Select the Blend workflow.
Find a PNG of a tree with a transparent background, then upload it or drag it onto the canvas.
Add an image of an architectural sketch the same way.
Select the sketch and send it behind the tree with Ctrl + B.
Resize and position the images.
Check whether the dotted canvas shows through inside the frame. If it does, add a background image, such as a landscape or plain white, and send it to the bottom.
Enter a positive and a negative prompt, then click Generate.
Click Clear, adjust the composition and generate again.
Compare your results in the Outputs tab.

A Blend collage of a tree in front of a house sketch, with the positive prompt “A mid century modern residential house” — **Fig. 15**A Blend composition: a tree over a house sketch, with the prompt “A mid century modern residential house”.

06Advanced parameters

Each workflow has its own set of parameters, and some are shared between workflows. The table summarises them; the sections that follow show their effect.

Parameter	Workflows	Default	What it controls
Steps	All	`20`	Number of denoising iterations. More steps, more detail, longer wait.
CFG	All	`8.5`	How strictly the result follows the prompts.
Iterate	All	`off`	Keeps the seed fixed so you can make small changes to a result.
Relative output height, width	Blend, Canny, Depth	`1`	Aspect ratio of the workspace.
Denoise	Blend	`0.80`	How much of the input collage is replaced.
ControlNet strength	Canny, Depth	–	How closely the result follows the detected edges or depth.
ControlNet start, end	Canny, Depth	–	When the conditioning applies during generation.

6.1Steps

Generation starts from random noise and removes it over a number of steps. Higher step counts give more resolved images but take longer. Very low values leave the image as noise.

Four generations of an architectural sketch with the same seed: the first is noise, later ones are increasingly resolved — **Fig. 16**Same seed, different step counts. From left to right: 5, 10, 20, 40. Prompt: “Architectural sketch”.

6.2CFG

The classifier-free guidance (CFG) scale controls how closely the generation follows the text prompts. A high CFG narrows the variety of results and can lower their quality. A low CFG gives images that follow the prompts only loosely.

Three generated brutalist buildings in a desert at increasing CFG — **Fig. 17**Same prompt, different CFG scales. From left to right: 1, 10, 20. Prompt: “Brutalist office building in the desert, photorealistic”.

6.3Iterate

The seed sets the random noise that each generation starts from. Tick Iterate to keep the seed the same between generations. You can then change one word of the prompt or one parameter and see exactly what that change does, without getting a completely new image.

6.4Relative output height and width

These sliders set the aspect ratio of the workspace in the Blend, Canny and Depth workflows. Adjust them to suit your composition, for example a wide frame for a street elevation.

6.5Denoise

In Blend, generation starts from your collage instead of from random noise. Denoise sets how much of the collage is replaced. Higher values keep less of the original composition, but still blend the inputs into one image that carries their concepts.

A collage of a sketched building, a tree and a figure, followed by three results that depart further from it — **Fig. 18**Same seed and prompt, different denoise values. From left to right: input, 0.60, 0.75, 1.00.

6.6ControlNet strength, start and end

Canny and Depth use ControlNets to condition the generation on the edges or depth of your input. Strength sets how closely the result follows that conditioning. Lower values give less resemblance; higher values give less variety. At 0.00 the input is ignored.

A Rubik’s cube input, followed by a round tower that ignores it and two cubic buildings that follow it — **Fig. 19**Same input, different ControlNet strength values. From left to right: input, 0.00, 0.50, 1.00.

Start and end set when the conditioning applies, as a fraction of the generation steps.

A start above 0 begins the conditioning later. The first steps run freely, which adds variety to the overall form.
An end below 1 stops the conditioning early. The last steps run freely, which adds variety to the details.

A Rubik’s cube input, followed by three cubic buildings with different start and end values — **Fig. 20**Same input, different ControlNet start and end values. From left to right: input, 0.15/1.00, 0.00/1.00, 0.00/0.15.

07Tips and troubleshooting

If	Try
Dark patches appear in the result	Part of the frame is empty. Add a background image and send it to the bottom.
The result ignores your collage	Lower Denoise in Blend, or raise ControlNet strength in Canny and Depth.
The result copies the input too closely	Raise Denoise, lower ControlNet strength, or end the conditioning earlier.
The style is not what you asked for	Name a specific style, medium, artist or building in the prompt. Generic architectural terms rarely work as well.
Each attempt looks completely different	Tick Iterate to keep the seed fixed while you adjust.
Generation takes long	Lower the step count while you explore, and raise it for the final image.