INITIALIZING BUNKROS IDENTITY LAB
LOC UNDERGROUND
SYS --:--:--
Learning / Image Generation

From noise to picture.

Image models don't "draw." They start with a screen of pure static and, step by step, remove the noise until a picture your words describe emerges. Once you can see that happening, prompting stops being magic and becomes direction. Slide the denoiser yourself and watch chaos resolve into an image.

You'll be able to

Explain diffusion

Describe how noise becomes a controlled image.

You'll be able to

Prompt with grammar

Direct subject, medium, light and style — not just nouns.

Watch for

"Pretty but wrong"

A gorgeous image can still miss the brief, text or rights.

01 — Warm-up

True or false?

"An AI image model paints a picture stroke by stroke, like a human artist starting from a blank canvas." True or false?

Self-Check · True / False
02 — Foundation

How diffusion works

During training, the model is shown millions of images that have been progressively destroyed with random noise, and it learns to reverse that destruction. At generation time it does exactly that in reverse: start from pure noise, then denoise step by step toward the most likely image for your prompt.

Core insight

The model never invents from a blank page — it sculpts noise away. A vague prompt gives it a vague target, so it denoises toward a generic average. A specific prompt gives it a sharp target. Vague in, mush out.

03 — Under the hood

What the model sees

Text encoder

Words → vector

A text encoder (CLIP/T5) turns your prompt into numbers that steer the denoising toward matching content.

Latent space

Compressed canvas

Most models work in a small compressed space, not raw pixels — which is why they're fast enough to run at all.

Guidance

How literally it obeys

A "guidance" (CFG) setting controls how strictly the model follows the prompt. Too high and the image "burns"; too low and it drifts.

Seed

The starting noise

The random seed is the exact static it starts from. Same seed + prompt + settings ≈ the same image.

04 — Interactive

The denoiser

Drag Denoising steps from 0 to 100 and watch pure static resolve into an image. Push Guidance up to make the target more vivid (and, too far, over-cooked). Hit Generate to animate the full reverse-diffusion, or Re-seed for a fresh starting noise.

Sandbox · Sculpt the noise away

At 0 steps it's pure noise — the model's blank slate. Slide right to denoise toward the prompt's target.

05 — Craft

Prompt craft for images

An image prompt is a brief, not a wish. Specify more than the subject — describe the picture as if briefing a photographer or illustrator.

Anatomy

Six layers of an image prompt

1 · Subject — who/what, with the details that matter.
2 · Composition — close-up, wide, low angle, centred.
3 · Medium — photo, 3D render, oil painting, line art.
4 · Lighting — golden hour, rim light, softbox, neon.
5 · Style & colour — mood, palette, era, technique (not a living artist's name).
6 · Parameters — aspect ratio, and what to avoid (negatives).

Template: [subject], [composition], [medium], [lighting], [style/colour] — [aspect] — no [unwanted]
Weak

"a cool poster"

No framing, light, medium or mood — the model fills every gap with the statistical average. Generic, forgettable.

Strong

The same idea, directed

"Riso-print gig poster for a queer synth night, bold halftone texture, acid-green and magenta, centred title space, 2:3 — no faces, no clutter."

Why it wins: every layer set, so you can change one variable per round.
Pro technique

Change one variable per iteration and lock the seed to hold composition while you tune wording. Use the negative prompt to steer away from artefacts (extra fingers, text, watermarks).

06 — In the Wild

Riso the zine cover

Devon makes covers for a community zine and wants a consistent risograph look across six issues.

Before

"make a cool zine cover"

Six wildly different styles — nothing reads as one series. The model averaged "cover" into noise.

After

Locked look + seed

Devon writes a full six-layer prompt, locks a style reference and a seed, and changes only the subject each issue.

Result: six covers that obviously belong together — same palette, grain and framing.
Why it matters · rights

Consistency comes from references and seeds, not luck. And before selling the zine, Devon checks the model's licence allows commercial use and avoids imitating a named living artist.

07 — Research

Map the tools

Image models change monthly. Find which one fits a specific need, and note one control it offers.

Search for

What to look up

"best AI image model for text in image" · "open-source image model local" · "commercially safe AI image generator".

Judge the source: prefer the maker's own docs and recent comparisons; check the date.
Your findings
Saved locally on your device.
08 — Assignment

Direct a poster

Write a single image prompt for a community-event poster using all six layers — subject, composition, medium, lighting, style/colour, and parameters/negatives.

Your image prompt
Saved locally.
Success rubric — tap to expand

Needs work: a bare subject ("a party poster").
Getting there: subject + one or two layers.
Solid: most layers present, but no negatives or parameters.
Excellent: all six layers, with a clear medium, lighting and a negative — ready to iterate one variable at a time.

09 — Retrieval practice

Quiz — 5 questions

Question 1 · Easy
How does a diffusion model create an image?
A
It paints stroke by stroke from a blank canvas.
B
It starts from random noise and removes it step by step toward the prompt.
C
It copy-pastes the closest image from its training set.
Question 2 · Medium
Why lock the "seed" while tuning a prompt?
A
It makes the image higher resolution.
B
It removes the need for a prompt.
C
It holds the composition steady so you can change one word at a time.
Question 3 · Easy
What is a "negative prompt"?
A
A list of things to steer the image away from.
B
A complaint you send the developers.
C
A prompt written in lowercase.
Question 4 · Medium · True / False
"Higher guidance is always better — set it to maximum." True or False?
T
True
F
False
Question 5 · Easy
Why does "a cool poster" produce a generic result?
A
The model is broken.
B
"Cool" is banned by safety filters.
C
A vague prompt is a vague target, so it denoises toward the statistical average.
10 — Avoid these

Common mistakes

Trap

One-shot perfectionism

Expecting the first image to ship. Fix: generate several, lock a seed, change one variable per round.

Trap

Prompt soup

40 contradictory adjectives = mush. Fix: name the artefact and medium clearly; cut noise words.

Trap

Artist-name reliance

"In the style of [living artist]" is unreliable and ethically/legally fraught. Fix: describe the technique instead.

Trap

"Pretty = done"

A striking image can still fail on text, hands, brand fit or rights. Fix: check usability and licence before shipping.

Debug your thinking
11 — Reflect

Direction, not dice

Now that the model sculpts noise toward your words, where does the authorship live — in the prompt, the curation, or the iteration? What would you never want a model to generate in your name?

Private journal
Saved locally to your device.
Well done

You can see the noise behind the picture now. Next, we move from single images to orchestrating whole creative pipelines.

12 — Reference

Glossary — precise words

Diffusion
Generating an image by starting from noise and removing it step by step toward the prompt.
Denoising step
One pass of removing noise; more steps = more refinement, up to a point.
Latent space
A compressed space where most models work instead of raw pixels, for speed.
Text encoder
The component (CLIP/T5) turning your words into a vector that steers generation.
Seed
The starting noise. Same seed + prompt + settings ≈ the same image.
Guidance (CFG)
How strictly the model follows the prompt. Too high "burns"; too low drifts.
Negative prompt
A list of traits or artefacts to steer the image away from.
Style reference
An image or trained add-on that locks a consistent look across outputs.
Inpainting
Regenerating only a masked region of an image.
Upscaling
Increasing resolution and detail after generation.
Aspect ratio
The width-to-height shape of the output (1:1, 16:9, 2:3).
Commercial licence
Terms permitting paid use of outputs; some tools also indemnify you.