A woman in an evening gown stands atop a city roof overlooking a sprawling metropolis, cinematic composition

--CONTROL BOX--

<lora:Age_v1:-1>

<lora:Elaborate_Dress50_Cake50_V1:3>

<lora:Facing_v1:1.50>

<lora:Illuminate_v1:-1>

<lora:Size_v1:0.0>

<lora:Transparent_v3:-1>

<lora:Weathered_v2:3>

<lora:Bootiful_Detroit_Slider:3>

<lora:Hirsute_v1:1>

Build a “Control Box” with Concept Sliders

Prompting is excellent for describing what you want in an image. It is often much less convenient when you want to make small, controlled adjustments to an image you already like.

Suppose you've generated a character and the composition is almost right, but you want her a little older, slightly more elaborately dressed, facing farther away from the camera, or the environment more weathered.

You can rewrite the prompt:

“slightly older woman, facing mostly away, moderately elaborate clothing…”

But changing prompt language can change much more than the attribute you intended to adjust. Composition, pose, clothing, lighting, and even the identity of the subject can move along with it.

Concept Sliders offer another approach: leave the descriptive prompt alone and put the adjustable properties into a separate set of LoRA controls.

The Concept Sliders researchers describe these LoRAs as low-rank adapters representing directions for particular attributes. At inference, changing the slider weight changes the strength or direction of that concept. They also specifically designed sliders to be composable: their experiments demonstrate compositions containing up to 50 sliders simultaneously.

Bau Lab — Concept Sliders

The Control Box

In an interface such as Forge, your prompt might normally look something like:

A woman standing on a rooftop overlooking a futuristic city at night,
long coat, dramatic city lights, cinematic photograph

Now add your Concept Slider LoRAs together in one convenient block:

A woman standing on a rooftop overlooking a futuristic city at night,
long coat, dramatic city lights, cinematic photograph

--- CONTROL BOX ---
<lora:ChronoShift_XL:0>
<lora:Facing_XL:0>
<lora:Elaborate_XL:0>
<lora:Weathered_XL:0>
<lora:Illuminate_XL:0>

Think of everything above the Control Box as the description of the picture.

Think of the Control Box as the control panel for the picture.

Instead of repeatedly rewriting adjectives in your prompt, you can adjust numbers:

--- CONTROL BOX ---
<lora:ChronoShift_XL:0.5>
<lora:Facing_XL:-1.5>
<lora:Elaborate_XL:2>
<lora:Weathered_XL:1>
<lora:Illuminate_XL:-0.5>

Depending on the polarity of the particular sliders, those numbers might mean: make the person older, turn her away from the viewer, make the design more elaborate, weather the environment, and darken the illumination.

The exact useful range varies from slider to slider. Zero is your neutral position. Positive and negative values move in opposite conceptual directions.

And zero is particularly useful because you don't have to keep adding and deleting LoRA tags. Your Control Box can simply stay in the prompt.

Start at Zero

When building a new image, start with the controls in neutral:

--- CONTROL BOX ---
<lora:ChronoShift_XL:0>
<lora:Facing_XL:0>
<lora:Elaborate_XL:0>
<lora:Weathered_XL:0>
<lora:Illuminate_XL:0>

Find a prompt, seed and composition you like first.

Then lock the seed.

This is important.

A random seed asks the diffusion model to construct another image every time you press Generate. That's useful while exploring, but terrible when you're trying to determine what one control actually does.

Once you've found an image worth developing, copy its seed into the seed field instead of leaving it at -1.

Now you have an experimental control.

Adjust One Knob at a Time

Suppose we want to change which way our rooftop character faces.

Keep:

• the same prompt,

• the same seed,

• the same checkpoint,

• the same sampler and scheduler,

• the same CFG,

• the same resolution,

• and every other slider at the same weight.

Change only Facing.

For example:

Facing: 0

Generate.

Then:

Facing: -0.5

Generate again.

Then:

Facing: -1

Then perhaps:

Facing: -1.5

You are no longer asking, “Which image do I like better?”

You're asking something much more useful:

What changed when I moved this one control?

If 0 looks toward the viewer, -0.5 turns partly away, -1 gives a stronger turn, and -1.5 gives you the rear-facing composition you wanted, you've found the setting for this image.

Leave it there.

Now adjust the next control.

Facing: -1.5
Elaborate: 0

becomes:

Facing: -1.5
Elaborate: 0.5

then perhaps:

Facing: -1.5
Elaborate: 1

You've now preserved the orientation you chose while searching for the amount of elaboration you want.

Repeat this process for each slider.

Don't Expect Every Slider to Behave Like a Photoshop Dial

This is an important distinction.

A Concept Slider weight is continuous, but the visual solutions produced by a diffusion checkpoint may not change continuously.

For something like Facing, for example, you shouldn't necessarily expect:

0      = exactly 0°
-0.5   = exactly 45°
-1     = exactly 90°
-1.5   = exactly 135°
-2     = exactly 180°

That's not what the slider knows.

It applies semantic pressure in a direction, and the checkpoint finds a coherent image satisfying that pressure.

So you may see something more like:

0       front
-0.5    three-quarter
-1      profile
-1.5    still profile, but different posture
-2      back

Another seed or pose may make the transition at different weights.

That's why fixed-seed sweeps are so useful. You're discovering how that particular slider interacts with that particular image rather than assuming that a number has a universal visual meaning.

Build the Image Like a Mixing Board

Once you've found useful settings individually, your Control Box might eventually look like:

--- CONTROL BOX ---
<lora:ChronoShift_XL:0.4>
<lora:Hirsute_XL:0.3>
<lora:Thick_XL:0.6>
<lora:Facing_XL:-1.75>
<lora:Elaborate_XL:1.5>
<lora:Weathered_XL:2>
<lora:Illuminate_XL:-0.75>

At that point the analogy isn't really a prompt anymore.

It's a mixing board.

Your ordinary prompt establishes the subject and scene. The sliders become knobs controlling selected characteristics of the model's solution.

And you can always go back.

If the character is perfect but the environment has become too decrepit:

Weathered: 2

becomes:

Weathered: 1

If the subject is facing too far away:

Facing: -1.75

becomes:

Facing: -1.25

You don't need to dismantle the prompt that produced the image in the first place.

Sliders Can Interact — So Test Them

“Composable” does not mean that every independently trained LoRA is mathematically guaranteed never to interact with every other LoRA.

The Bau Lab work demonstrates that their low-rank Concept Sliders can be composed at considerable scale—up to 50 sliders in their qualitative demonstration—and describes composability as one of the method's advantages.

But in practical community workflows, sliders may come from different trainers, datasets, base models and training methods.

So use the Control Box method itself to catch interactions.

If your image looks great with:

Facing: -1.5
Elaborate: 1
Weathered: 1

and suddenly changes strangely when you add:

ChronoShift: 0.75

don't change five things trying to fix it.

Set ChronoShift back to zero.

Confirm the image returns.

Then sweep ChronoShift alone.

You've just turned ordinary image generation into a tiny controlled experiment.

A Simple Workflow

1. Write the image prompt. Get the subject, scene and composition approximately where you want them.

2. Choose a good seed and lock it. Don't continue with -1.

3. Put your sliders into a Control Box. Start unfamiliar controls at 0.

4. Choose one control. Change only its weight while everything else stays fixed.

5. Generate a small sweep. For example -1, -0.5, 0, +0.5, +1, adjusted for the slider's recommended range.

6. Pick the result you prefer. Keep that weight.

7. Move to the next slider. Repeat.

8. Once the image is dialed in, unlock the seed if you want variations. Your Control Box now describes the conceptual recipe you developed.

The important habit is simple:

One knob. Same seed. Compare. Choose. Next knob.

That's the Control Box.