We're opening up Maki through a series of Medium posts. We'll describe what Maki is building and how. We'll explain how we built a SaaS in 100 days. This isn't a step-by-step guide to starting a business, but a product philosophy to help you build quickly.

Focus on end-users.

This may seem obvious, but it's important to talk to users first. When we started thinking about Maki, we interviewed 100 HR directors from SMBs, Next40 scale-ups, and major Enterprises to understand their concerns. We initially thought about the future of employment, but most of their worries centered on Recruitment. We don't have a clear recruiting process, it's time-consuming, we recruit clones, we don't support diversity, etc. And as hiring managers, we couldn't help but agree.

Co-create your product with your end-users.

We went to the drawing board, read as many books as possible (here, here, and here), and when we started getting a sense for a solution, we questioned 100 more operational HR specialists to corroborate the idea and get a feel for our potential answer. This confirmed our direction to help hire more objectively and efficiently.

Back to the drawing board, we designed our first flows and screens. We organized sessions with certain survey respondents to show them our early work and get comments. We got great input that helped us build Maki, and we met some consumers. Obsess about users and execute alongside them.

Don’t shoot for the moon, yet. Make pragmatic choices first.

Once we were convinced, we began building. To launch a SaaS in 100 days, we needed an operating principle that allowed us to accelerate while still providing a reliable, secure, scalable experience. We focused on adding value and outsourced everything else. Example:

Concentrate on adding value. Reuse existing bricks.

When determining which technology to use, we looked at our strengths and the future to see what would last. Node.js for backend, React for frontend, both with typescript. We thought this technique would scale well since it would attract more talent and the surrounding mature ecosystem would help us go quicker.

We explored for ways to bootstrap services while setting down strong foundations that might support millions of users. We built our backend services on NestJS so we could extend into microservices later. Hasura, a GraphQL APIs engine, automates Postgres data exposing through a graphQL layer. MUI's ready-to-use components powered our design-system. We used well-maintained open-source projects to speed up certain tasks.

We outsourced important components of our platform (Auth0 for authentication, Stripe for billing, SendGrid for notifications) because, let's face it, we couldn't do better. We choose to host our complete infrastructure (SQL, Cloud run, Logs, Monitoring) on GCP to simplify our work between numerous providers.

Focus on your business, use existing bricks for the rest. For the curious, we'll shortly publish articles detailing each stage.

Most importantly, empower people and step back.

We couldn't have done this without the incredible people who have supported us from the start. Since Powership is one of our key values, we provided our staff the power to make autonomous decisions from day one. Because we believe our firm is its people, we hired smart builders and let them build.

Nicolas left Spendesk to create scalable interfaces using react-router, react-queries, and MUI. JD joined Swile and chose Hasura as our GraphQL engine. Jérôme chose NestJS to build our backend services. Since then, Justin, Ben, Anas, Yann, Benoit, and others have followed suit.

If you consider your team a collective brain, you should let them make decisions instead of directing them what to do. You'll make mistakes, but you'll go faster and learn faster overall.

Invest in great talent and develop a strong culture from the start. Here's how to establish a SaaS in 100 days.

His success was unforeseeable.

2014 saw Facebook's $2.3 billion purchase of Oculus VR.

19-year-old Palmer Luckey founded Oculus. He quit journalism school. His parents worried about his college dropout.

Facebook bought Oculus VR in less than 2 years.

Palmer Luckey started Anduril Industries. Palmer has raised $385 million with Anduril.

The Oculus journey began in a trailer

Palmer Luckey, 19, owned the trailer.

Luckey had his trailer customized. The trailer had all six of Luckey's screens. In the trailer's remaining area, Luckey conducted hardware tests.

At 16, he became obsessed with virtual reality. Virtual reality was rare at the time.

Luckey didn't know about VR when he started.

Previously, he liked "portabilizing" mods. Hacking ancient game consoles into handhelds.

In his city, fewer portabilizers actively traded.

Luckey started "ModRetro" for other portabilizers. Luckey was exposed to VR headsets online.

Luckey:

“Man, ModRetro days were the best.”

Palmer Luckey used VR headsets for three years. His design had 50 prototypes.

Luckey used to work at the Long Beach Sailing Center for minimum salary, servicing diesel engines and cleaning boats.

Luckey worked in a USC Institute for Creative Technologies mixed reality lab in July 2011. (ICT).

Luckey cleaned the lab, did reports, and helped other students with VR projects.

Luckey's lab job was dull.

Luckey chose to work in the lab because he wanted to engage with like-minded folks.

By 2012, Luckey had a prototype he hoped to share globally. He made cheaper headsets than others.

Luckey wanted to sell an easy-to-assemble virtual reality kit on Kickstarter.

He realized he needed a corporation to do these sales legally. He started looking for names. "Virtuality," "virtual," and "VR" are all taken.

Hence, Oculus.

If Luckey sold a hundred prototypes, he would be thrilled since it would boost his future possibilities.

John Carmack, legendary game designer

Carmack has liked sci-fi and fantasy since infancy.

Carmack loved imagining intricate gaming worlds.

His interest in programming and computer science grew with age.

He liked graphics. He liked how mismatching 0 and 1 might create new colors and visuals.

Carmack played computer games as a teen. He created Shadowforge in high school.

He founded Id software in 1991. When Carmack created id software, console games were the best-sellers.

Old computer games have weak graphics. John Carmack and id software developed "adaptive tile refresh."

This technique smoothed PC game scrolling. id software launched 3-D, Quake, and Doom using "adaptive tile refresh."

These games made John Carmack a gaming star. Later, he sold Id software to ZeniMax Media.

How Palmer Luckey met Carmack

In 2011, Carmack was thinking a lot about 3-D space and virtual reality.

He was underwhelmed by the greatest HMD on the market. Because of their flimsiness and latency.

His disappointment was partly due to the view (FOV). Best HMD had 40-degree field of view.

Poor. The best VR headset is useless with a 40-degree FOV.

Carmack intended to show the press Doom 3 in VR. He explored VR headsets and internet groups for this reason.

Carmack identified a VR enthusiast in the comments section of "LEEP on the Cheap." "PalmerTech" was the name.

Carmack approached PalmerTech about his prototype. He told Luckey about his VR demos, so he wanted to see his prototype.

Carmack got a Rift prototype. Here's his May 17 tweet.

John Carmack tweeted an evaluation of the Luckey prototype.

Dan Newell, a Valve engineer, and Mick Hocking, a Sony senior director, pre-ordered Oculus Rift prototypes with Carmack's help.

Everyone praised Luckey after Carmack demoed Rift.

Palmer Luckey received a job offer from Sony.

• It was a full-time position at Sony Computer Europe.

• He would run Sony’s R&D lab.

• The salary would be $70k.

Who is Brendan Iribe?

Brendan Iribe started early with Startups. In 2004, he and Mike Antonov founded Scaleform.

Scaleform created high-performance middleware. This package allows 3D Flash games.

In 2011, Iribe sold Scaleform to Autodesk for $36 million.

How Brendan Iribe discovered Palmer Luckey.

Brendan Iribe's friend Laurent Scallie.

Laurent told Iribe about a potential opportunity.

Laurent promised Iribe VR will work this time. Laurent introduced Iribe to Luckey.

Iribe was doubtful after hearing Laurent's statements. He doubted Laurent's VR claims.

But since Laurent took the name John Carmack, Iribe thought he should look at Luckey Innovation. Iribe was hooked on virtual reality after reading Palmer Luckey stories.

He asked Scallie about Palmer Luckey.

Iribe convinced Luckey to start Oculus with him

First meeting between Palmer Luckey and Iribe.

The Iribe team wanted Luckey to feel comfortable.

Iribe sought to convince Luckey that launching a company was easy. Iribe told Luckey anyone could start a business.

Luckey told Iribe's staff he was homeschooled from childhood. Luckey took self-study courses.

Luckey had planned to launch a Kickstarter campaign and sell kits for his prototype. Many companies offered him jobs, nevertheless.

He's considering Sony's offer.

Iribe advised Luckey to stay independent and not join a firm. Iribe asked Luckey how he could raise his child better. No one sees your baby like you do?

Iribe's team pushed Luckey to stay independent and establish a software ecosystem around his device.

After conversing with Iribe, Luckey rejected every job offer and merger option.

Iribe convinced Luckey to provide an SDK for Oculus developers.

After a few months. Brendan Iribe co-founded Oculus with Palmer Luckey. Luckey trusted Iribe and his crew, so he started a corporation with him.

Crowdfunding

Brendan Iribe and Palmer Luckey launched a Kickstarter.

Gabe Newell endorsed Palmer's Kickstarter video.

Gabe Newell wants folks to trust Palmer Luckey since he's doing something fascinating and answering tough questions.

Mark Bolas and David Helgason backed Palmer Luckey's VR Kickstarter video.

Luckey introduced Oculus Rift during the Kickstarter campaign. He introduced virtual reality during press conferences.

Oculus' Kickstarter effort was a success. Palmer Luckey felt he could raise $250,000.

Oculus raised $2.4 million through Kickstarter. Palmer Luckey's virtual reality vision was well-received.

Mark Zuckerberg's Oculus discovery

Brendan Iribe and Palmer Luckey hired the right personnel after a successful Kickstarter campaign.

Oculus needs a lot of money for engineers and hardware. They needed investors' money.

Series A raised $16M.

Next, Andreessen Horowitz partner Brain Cho approached Iribe.

Cho told Iribe that Andreessen Horowitz could invest in Oculus Series B if the company solved motion sickness.

Mark Andreessen was Iribe's dream client.

Marc Andreessen and his partners gave Oculus $75 million.

Andreessen introduced Iribe to Zukerberg. Iribe and Zukerberg discussed the future of games and virtual reality by phone.

Facebook's Oculus demo

Iribe showed Zuckerberg Oculus.

Mark was hooked after using Oculus. The headset impressed him.

The whole Facebook crew who saw the demo said only one thing.

“Holy Crap!”

This surprised them all.

Mark Zuckerberg was impressed by the team's response. Mark Zuckerberg met the Oculus team five days after the demo.

First meeting Palmer Luckey.

Palmer Luckey is one of Mark's biggest supporters and loves Facebook.

Oculus Acquisition

Zuckerberg wanted Oculus.

Brendan Iribe had requested for $4 billion, but Mark wasn't interested.

Facebook bought Oculus for $2.3 billion after months of drama.

After selling his company, how does Palmer view money?

Palmer loves the freedom money gives him. Money frees him from small worries.

Money has allowed him to pursue things he wouldn't have otherwise.

“If I didn’t have money I wouldn’t have a collection of vintage military vehicles…You can have nice hobbies that keep you relaxed when you have money.”

He didn't start Oculus to generate money. His virtual reality passion spanned years.

He didn't have to lie about how virtual reality will transform everything until he needed funding.

The company's success was an unexpected bonus. He was merely passionate about a good cause.

After Oculus' $2.3 billion exit, what changed?

Palmer didn't mind being rich. He did similar things.

After Facebook bought Oculus, he moved to Silicon Valley and lived in a 12-person shared house due to high rents.

Palmer might have afforded a big mansion, but he prefers stability and doing things because he wants to, not because he has to.

“Taco Bell is never tasted so good as when you know you could afford to never eat taco bell again.”

Palmer's leadership shifted.

Palmer changed his leadership after selling Oculus.

When he launched his second company, he couldn't work on his passions.

“When you start a tech company you do it because you want to work on a technology, that is why you are interested in that space in the first place. As the company has grown, he has realized that if he is still doing optical design in the company it’s because he is being negligent about the hiring process.”

Once his startup grows, the founder's responsibilities shift. He must recruit better firm managers.

Recruiting talented people becomes the top priority. The founder must convince others of their influence.

A book that helped me write this:

The History of the Future: Oculus, Facebook, and the Revolution That Swept Virtual Reality — Blake Harris

*This post is a summary. Read the full article here.

Niches are rich, as the adage goes.

Micro SaaS requires a great micro-niche; otherwise, it's merely plain old SaaS with a large audience.

Instead of targeting broad markets with few identifying qualities, specialise down to a micro-niche. How would you target these users?

Better go tiny. You'll locate and engage new consumers more readily and serve them better with a customized solution.

Imagine you're a real estate lawyer looking for a case management solution. Because it's so specific to you, you'd be lured to this link:

instead of below:

Next, locate mini SaaS niches that could work for you. You're not yet looking at the problems/solutions in these areas, merely shortlisting them.

The market should be growing, not shrinking

We shouldn't design apps for a declining niche. We intend to target stable or growing niches for the next 5 to 10 years.

If it's a developing market, you may be able to claim a stake early. You must balance this strategy with safer, longer-established niches (accountancy, law, health, etc).

First Micro SaaS apps I designed were for Merch By Amazon creators, a burgeoning niche. I found this niche when searching for passive income.

Graphic designers and entrepreneurs post their art to Amazon to sell on clothes. When Amazon sells their design, they get a royalty. Since 2015, this platform and specialty have grown dramatically.

Amazon doesn't publicize the amount of creators on the platform, but it's possible to approximate by looking at Facebook groups, Reddit channels, etc.

I could see the community growing week by week, with new members joining. Merch was an up-and-coming niche, and designers made money when their designs sold. All I had to do was create tools that let designers focus on making bestselling designs.

Look at the Google Trends graph below to see how this niche has evolved and when I released my apps and resigned my job.

Are the users able to afford the tools?

Who's your average user? Consumer or business? Is your solution budgeted?

If they're students, you'll struggle to convince them to subscribe to your study-system app (ahead of video games and beer).

Let's imagine you designed a Shopify plugin that emails customers when a product is restocked. If your plugin just needs 5 product sales a month to justify its cost, everyone wins (just be mindful that one day Shopify could potentially re-create your plugins functionality within its core offering making your app redundant ).

Do specialized users buy tools? If so, that's comforting. If not, you'd better have a compelling value proposition for your end customer if you're the first.

This should include how much time or money your program can save or make the user.

Are you able to understand the Micro SaaS market?

Ideally, you're already familiar about the industry/niche. Maybe you're fixing a challenge from your day job or freelance work.

If not, evaluate how long it would take to learn the niche's users. Health & Fitness is easier to relate to and understand than hedge fund derivatives trading.

Competing in these complex (and profitable) fields might offer you an edge.

B2C, B2M, or B2B?

Consider your user base's demographics. Will you target businesses, consumers, or both? Let's examine the different consumer types:

• B2B refers to business-to-business transactions where customers are other businesses. UpVoty, Plutio, Slingshot, Salesforce, Atlassian, and Hubspot are a few examples of SaaS, ranging from Micro SaaS to SaaS.

• Business to Consumer (B2C), in which your clients are people who buy things. For instance, Duolingo, Canva, and Nomad List.

• For instance, my tool KDP Wizard has a mixed user base of publishing enterprises and also entrepreneurial consumers selling low-content books on Amazon. This is a case of business to many (B2M), where your users are a mixture of businesses and consumers. There is a large SaaS called Dropbox that offers both personal and business plans.

Targeting a B2B vs. B2C niche is very different. The sales cycle differs.

• A B2B sales staff must make cold calls to potential clients' companies. Long sales, legal, and contractual conversations are typically required for each business to get the go-ahead. The cost of obtaining a new customer is substantially more than it is for B2C, despite the fact that the recurring fees are significantly higher.

• Since there is typically only one individual making the purchasing decision, B2C signups are virtually always self-service with reduced recurring fees. Since there is typically no outbound sales staff in B2C, acquisition costs are significantly lower than in B2B.

User Characteristics for B2B vs. B2C

Consider where your niche's users congregate if you don't already have a presence there.

B2B users frequent LinkedIn and Twitter. B2C users are on Facebook/Instagram/Reddit/Twitter, etc.

Churn is higher in B2C because consumers haven't gone through all the hoops of a B2B sale. Consumers are more unpredictable than businesses since they let their bank cards exceed limitations or don't update them when they expire.

With a B2B solution, there's a contractual arrangement and the firm will pay the subscription as long as they need it.

Depending on how you feel about the above (sales team vs. income vs. churn vs. targeting), you'll know which niches to pursue.

You ought to respect potential customers.

Would you hang out with customers?

You'll connect with users at conferences (in-person or virtual), webinars, seminars, screenshares, Facebook groups, emails, support calls, support tickets, etc.

If talking to a niche's user base makes you shudder, you're in for a tough road. Whether they're demanding or dull, avoid them if possible.

Merch users are mostly graphic designers, side hustlers, and entrepreneurs. These laid-back users embrace technologies that assist develop their Merch business.

I discovered there was only one annual conference for this specialty, held in Seattle, USA. I decided to organize a conference for UK/European Merch designers, despite never having done so before.

Hosting a conference for over 80 people was stressful, and it turned out to be much bigger than expected, with attendees from the US, Europe, and the UK.

I met many specialized users, built relationships, gained trust, and picked their brains in person. Many of the attendees were already Merch Wizard users, so hearing their feedback and ideas for future features was invaluable.

focused and specific

Instead of building for a generic, hard-to-reach market, target a specific group.

I liken it to fishing in a little, hidden pond. This small pond has only one species of fish, so you learn what bait it likes. Contrast that with trawling for hours to catch as many fish as possible, even if some aren't what you want.

In the case management scenario, it's difficult to target leads because several niches could use the app. Where do your potential customers hang out? Your generic solution: No.

It's easier to join a community of Real Estate Lawyers and see if your software can answer their pain points.

My Success with Micro SaaS

In my case, my Micro SaaS apps have been my chrome extensions. Since I launched them, they've earned me an average $10k MRR, allowing me to quit my lousy full-time job years ago.

I sold my apps after scaling them for a life-changing lump amount. Since then, I've helped unfulfilled software developers escape the 9-5 through Micro SaaS.

Whether it's a profitable side hustle or a liferaft to quit their job and become their own Micro SaaS boss.

Having built my apps to the point where I could quit my job, then scaled and sold them, I feel I can share my skills with software developers worldwide.

Read my free guide on self-funded SaaS to discover more about Micro SaaS, or download your own copy. 12 chapters cover everything from Idea to Exit.

Watch my YouTube video to learn how to construct a Micro SaaS app in 10 steps.

Why breakthrough technologies must be accessible

ChatGPT has exploded. Over 1 million people have used the app, and coding sites like Stack Overflow have banned its answers. It's huge.

I wouldn't have called that as an AI researcher. ChatGPT uses the same GPT-3 technology that's been around for over two years.

More than impressive technology, ChatGPT 3 shows how access makes breakthroughs usable. OpenAI has finally made people realize the power of AI by packaging GPT-3 for normal users.

We think of Thomas Edison as the inventor of the lightbulb, not because he invented it, but because he popularized it.

Going forward, AI companies that make using AI easy will thrive.

Use-case importance

Most modern AI systems use massive language models. These language models are trained on 6,000+ years of human text.

GPT-3 ate 8 billion pages, almost every book, and Wikipedia. It created an AI that can write sea shanties and solve coding problems.

Nothing new. I began beta testing GPT-3 in 2020, but the system's basics date back further.

Tools like GPT-3 are hidden in many apps. Many of the AI writing assistants on this platform are just wrappers around GPT-3.

Lots of online utilitarian text, like restaurant menu summaries or city guides, is written by AI systems like GPT-3. You've probably read GPT-3 without knowing it.

Accessibility

Why is ChatGPT so popular if the technology is old?

ChatGPT makes the technology accessible. Free to use, people can sign up and text with the chatbot daily. ChatGPT isn't revolutionary. It does it in a way normal people can access and be amazed by.

Accessibility isn't easy. OpenAI's Sam Altman tweeted that opening ChatGPT to the public increased computing costs.

Each chat costs "low-digit cents" to process. OpenAI probably spends several hundred thousand dollars a day to keep ChatGPT running, with no immediate business case.

Academic researchers and others who developed GPT-3 couldn't afford it. Without resources to make technology accessible, it can't be used.

Retrospective

This dynamic is old. In the history of science, a researcher with a breakthrough idea was often overshadowed by an entrepreneur or visionary who made it accessible to the public.

We think of Thomas Edison as the inventor of the lightbulb. But really, Vasilij Petrov, Thomas Wright, and Joseph Swan invented the lightbulb. Edison made technology visible and accessible by electrifying public buildings, building power plants, and wiring.

Edison probably lost a ton of money on stunts like building a power plant to light JP Morgan's home, the NYSE, and several newspaper headquarters.

People wanted electric lights once they saw their benefits. By making the technology accessible and visible, Edison unlocked a hugely profitable market.

Similar things are happening in AI. ChatGPT shows that developing breakthrough technology in the lab or on B2B servers won't change the culture.

AI must engage people's imaginations to become mainstream. Before the tech impacts the world, people must play with it and see its revolutionary power.

As the field evolves, companies that make the technology widely available, even at great cost, will succeed.

OpenAI's compute fees are eye-watering. Revolutions are costly.

Without self-control, you'll waste time chasing dopamine fixes.

I found a powerful quote in Tony Robbins' Awaken The Giant Within:

“Most of the challenges that we have in our personal lives come from a short-term focus” — Tony Robbins

Most people are short-term oriented, but highly successful people are long-term oriented.

Successful people act in line with their long-term goals and values, while the rest are distracted by short-term pleasures and dopamine fixes.

Instant gratification wrecks lives

Instant pleasure is fleeting. Quickly fading effects leave you craving more stimulation.

Before you know it, you're in a cycle of quick fixes. This explains binging on food, social media, and Netflix.

These things cause a dopamine spike, which is entertaining. This dopamine spike crashes quickly, leaving you craving more stimulation.

It's fine to watch TV or play video games occasionally. Problems arise when brain impulses aren't controlled. You waste hours chasing dopamine fixes.

Instant gratification becomes problematic when it interferes with long-term goals, happiness, and life fulfillment.

Most rewarding things require delay

Life's greatest rewards require patience and delayed gratification. They must be earned through patience, consistency, and effort.

Ex:

• A fit, healthy body

• A deep connection with your spouse

• A thriving career/business

• A healthy financial situation

These are some of life's most rewarding things, but they take work and patience. They all require the ability to delay gratification.

To have a healthy bank account, you must save (and invest) a large portion of your monthly income. This means no new tech or clothes.

If you want a fit, healthy body, you must eat better and exercise three times a week. So no fast food and Netflix.

It's a battle between what you want now and what you want most.

Successful people choose what they want most over what they want now. It's a major difference.

Instant vs. delayed gratification

Most people subconsciously prefer instant rewards over future rewards, even if the future rewards are more significant.

We humans aren't logical. Emotions and instincts drive us. So we act against our goals and values.

Fortunately, instant gratification bias can be overridden. This is a modern superpower. Effective methods include:

#1: Train your brain to handle overstimulation

Training your brain to function without constant stimulation is a powerful change. Boredom can lead to long-term rewards.

Unlike impulsive shopping, saving money is boring. Having lots of cash is amazing.

Compared to video games, deep work is boring. A successful online business is rewarding.

Reading books is boring compared to scrolling through funny videos on social media. Knowledge is invaluable.

You can't do these things if your brain is overstimulated. Your impulses will control you. To reduce overstimulation addiction, try:

• Daily meditation (10 minutes is enough)

• Daily study/work for 90 minutes (no distractions allowed)

• First hour of the day without phone, social media, and Netflix

• Nature walks, journaling, reading, sports, etc.

#2: Make Important Activities Less Intimidating

Instant gratification helps us cope with stress. Starting a book or business can be intimidating. Video games and social media offer a quick escape in such situations.

Make intimidating tasks less so. Break them down into small tasks. Start a new business/side-hustle by:

• Get domain name

• Design website

• Write out a business plan

• Research competition/peers

• Approach first potential client

Instead of one big mountain, divide it into smaller sub-tasks. This makes a task easier and less intimidating.

#3: Plan ahead for important activities

Distractions will invade unplanned time. Your time is dictated by your impulses, which are usually Netflix, social media, fast food, and video games. It wants quick rewards and dopamine fixes.

Plan your days and be proactive with your time. Studies show that scheduling activities makes you 3x more likely to do them.

To achieve big goals, you must plan. Don't gamble.

Want to get fit? Schedule next week's workouts. Want a side-job? Schedule your work time.

Deep learning-based generative art is being researched. As usual, self-learning is better. Some models, like OpenAI's DALL-E 2, require registration and can only be used online, but others can be used locally, which is usually more enjoyable for curious users. I'll demonstrate the Stable Diffusion model's operation on a standard PC.

Let’s get started.

What It Does

Stable Diffusion uses numerous components:

• A generative model trained to produce images is called a diffusion model. The model is incrementally improving the starting data, which is only random noise. The model has an image, and while it is being trained, the reversed process is being used to add noise to the image. Being able to reverse this procedure and create images from noise is where the true magic is (more details and samples can be found in the paper).

• An internal compressed representation of a latent diffusion model, which may be altered to produce the desired images, is used (more details can be found in the paper). The capacity to fine-tune the generation process is essential because producing pictures at random is not very attractive (as we can see, for instance, in Generative Adversarial Networks).

• A neural network model called CLIP (Contrastive Language-Image Pre-training) is used to translate natural language prompts into vector representations. This model, which was trained on 400,000,000 image-text pairs, enables the transformation of a text prompt into a latent space for the diffusion model in the scenario of stable diffusion (more details in that paper).

This figure shows all data flow:

The weights file size for Stable Diffusion model v1 is 4 GB and v2 is 5 GB, making the model quite huge. The v1 model was trained on 256x256 and 512x512 LAION-5B pictures on a 4,000 GPU cluster using over 150.000 NVIDIA A100 GPU hours. The open-source pre-trained model is helpful for us. And we will.

Install

Before utilizing the Python sources for Stable Diffusion v1 on GitHub, we must install Miniconda (assuming Git and Python are already installed):

wget https://repo.anaconda.com/miniconda/Miniconda3-py39_4.12.0-Linux-x86_64.sh
chmod +x Miniconda3-py39_4.12.0-Linux-x86_64.sh
./Miniconda3-py39_4.12.0-Linux-x86_64.sh
conda update -n base -c defaults conda

Install the source and prepare the environment:

git clone https://github.com/CompVis/stable-diffusion
cd stable-diffusion
conda env create -f environment.yaml
conda activate ldm
pip3 install transformers --upgrade

Download the pre-trained model weights next. HiggingFace has the newest checkpoint sd-v14.ckpt (a download is free but registration is required). Put the file in the project folder and have fun:

python3 scripts/txt2img.py --prompt "hello world" --plms --ckpt sd-v1-4.ckpt --skip_grid --n_samples 1

Almost. The installation is complete for happy users of current GPUs with 12 GB or more VRAM. RuntimeError: CUDA out of memory will occur otherwise. Two solutions exist.

Running the optimized version

Try optimizing first. After cloning the repository and enabling the environment (as previously), we can run the command:

python3 optimizedSD/optimized_txt2img.py --prompt "hello world" --ckpt sd-v1-4.ckpt --skip_grid --n_samples 1

Stable Diffusion worked on my visual card with 8 GB RAM (alas, I did not behave well enough to get NVIDIA A100 for Christmas, so 8 GB GPU is the maximum I have;).

Running Stable Diffusion without GPU

If the GPU does not have enough RAM or is not CUDA-compatible, running the code on a CPU will be 20x slower but better than nothing. This unauthorized CPU-only branch from GitHub is easiest to obtain. We may easily edit the source code to use the latest version. It's strange that a pull request for that was made six months ago and still hasn't been approved, as the changes are simple. Readers can finish in 5 minutes:

• Replace if attr.device!= torch.device(cuda) with if attr.device!= torch.device(cuda) and torch.cuda.is available at line 20 of ldm/models/diffusion/ddim.py ().

• Replace if attr.device!= torch.device(cuda) with if attr.device!= torch.device(cuda) and torch.cuda.is available in line 20 of ldm/models/diffusion/plms.py ().

• Replace device=cuda in lines 38, 55, 83, and 142 of ldm/modules/encoders/modules.py with device=cuda if torch.cuda.is available(), otherwise cpu.

• Replace model.cuda() in scripts/txt2img.py line 28 and scripts/img2img.py line 43 with if torch.cuda.is available(): model.cuda ().

Run the script again.

Testing

Test the model. Text-to-image is the first choice. Test the command line example again:

python3 scripts/txt2img.py --prompt "hello world" --plms --ckpt sd-v1-4.ckpt --skip_grid --n_samples 1

The slow generation takes 10 seconds on a GPU and 10 minutes on a CPU. Final image:

Hello world is dull and abstract. Try a brush-wielding hamster. Why? Because we can, and it's not as insane as Napoleon's cat. Another image:

Generating an image from a text prompt and another image is interesting. I made this picture in two minutes using the image editor (sorry, drawing wasn't my strong suit):

I can create an image from this drawing:

python3 scripts/img2img.py --prompt "A bird is sitting on a tree branch" --ckpt sd-v1-4.ckpt --init-img bird.png --strength 0.8

It was far better than my initial drawing:

I hope readers understand and experiment.

Stable Diffusion UI

Developers love the command line, but regular users may struggle. Stable Diffusion UI projects simplify image generation and installation. Simple usage:

• Unpack the ZIP after downloading it from https://github.com/cmdr2/stable-diffusion-ui/releases. Linux and Windows are compatible with Stable Diffusion UI (sorry for Mac users, but those machines are not well-suitable for heavy machine learning tasks anyway;).

• Start the script.

Done. The web browser UI makes configuring various Stable Diffusion features (upscaling, filtering, etc.) easy:

V2.1 of Stable Diffusion

I noticed the notification about releasing version 2.1 while writing this essay, and it was intriguing to test it. First, compare version 2 to version 1:

• alternative text encoding. The Contrastive LanguageImage Pre-training (CLIP) deep learning model, which was trained on a significant number of text-image pairs, is used in Stable Diffusion 1. The open-source CLIP implementation used in Stable Diffusion 2 is called OpenCLIP. It is difficult to determine whether there have been any technical advancements or if legal concerns were the main focus. However, because the training datasets for the two text encoders were different, the output results from V1 and V2 will differ for the identical text prompts.

• a new depth model that may be used to the output of image-to-image generation.

• a revolutionary upscaling technique that can quadruple the resolution of an image.

• Generally higher resolution Stable Diffusion 2 has the ability to produce both 512x512 and 768x768 pictures.

The Hugging Face website offers a free online demo of Stable Diffusion 2.1 for code testing. The process is the same as for version 1.4. Download a fresh version and activate the environment:

conda deactivate  
conda env remove -n ldm  # Use this if version 1 was previously installed
git clone https://github.com/Stability-AI/stablediffusion
cd stablediffusion
conda env create -f environment.yaml
conda activate ldm

Hugging Face offers a new weights ckpt file.

The Out of memory error prevented me from running this version on my 8 GB GPU. Version 2.1 fails on CPUs with the slow conv2d cpu not implemented for Half error (according to this GitHub issue, the CPU support for this algorithm and data type will not be added). The model can be modified from half to full precision (float16 instead of float32), however it doesn't make sense since v1 runs up to 10 minutes on the CPU and v2.1 should be much slower. The online demo results are visible. The same hamster painting with a brush prompt yielded this result:

It looks different from v1, but it functions and has a higher resolution.

The superresolution.py script can run the 4x Stable Diffusion upscaler locally (the x4-upscaler-ema.ckpt weights file should be in the same folder):

python3 scripts/gradio/superresolution.py configs/stable-diffusion/x4-upscaling.yaml x4-upscaler-ema.ckpt

This code allows the web browser UI to select the image to upscale:

The copy-paste strategy may explain why the upscaler needs a text prompt (and the Hugging Face code snippet does not have any text input as well). I got a GPU out of memory error again, although CUDA can be disabled like v1. However, processing an image for more than two hours is unlikely:

Stable Diffusion Limitations

When we use the model, it's fun to see what it can and can't do. Generative models produce abstract visuals but not photorealistic ones. This fundamentally limits The generative neural network was trained on text and image pairs, but humans have a lot of background knowledge about the world. The neural network model knows nothing. If someone asks me to draw a Chinese text, I can draw something that looks like Chinese but is actually gibberish because I never learnt it. Generative AI does too! Humans can learn new languages, but the Stable Diffusion AI model includes only language and image decoder brain components. For instance, the Stable Diffusion model will pull NO WAR banner-bearers like this:

V1:

V2.1:

The shot shows text, although the model never learned to read or write. The model's string tokenizer automatically converts letters to lowercase before generating the image, so typing NO WAR banner or no war banner is the same.

I can also ask the model to draw a gorgeous woman:

V1:

V2.1:

The first image is gorgeous but physically incorrect. A second one is better, although it has an Uncanny valley feel. BTW, v2 has a lifehack to add a negative prompt and define what we don't want on the image. Readers might try adding horrible anatomy to the gorgeous woman request.

If we ask for a cartoon attractive woman, the results are nice, but accuracy doesn't matter:

V1:

V2.1:

Another example: I ordered a model to sketch a mouse, which looks beautiful but has too many legs, ears, and fingers:

V1:

V2.1: improved but not perfect.

V1 produces a fun cartoon flying mouse if I want something more abstract:

I tried multiple times with V2.1 but only received this:

The image is OK, but the first version is closer to the request.

Stable Diffusion struggles to draw letters, fingers, etc. However, abstract images yield interesting outcomes. A rural landscape with a modern metropolis in the background turned out well:

V1:

V2.1:

Generative models help make paintings too (at least, abstract ones). I searched Google Image Search for modern art painting to see works by real artists, and this was the first image:

I typed "abstract oil painting of people dancing" and got this:

V1:

V2.1:

It's a different style, but I don't think the AI-generated graphics are worse than the human-drawn ones.

The AI model cannot think like humans. It thinks nothing. A stable diffusion model is a billion-parameter matrix trained on millions of text-image pairs. I input "robot is creating a picture with a pen" to create an image for this post. Humans understand requests immediately. I tried Stable Diffusion multiple times and got this:

This great artwork has a pen, robot, and sketch, however it was not asked. Maybe it was because the tokenizer deleted is and a words from a statement, but I tried other requests such robot painting picture with pen without success. It's harder to prompt a model than a person.

I hope Stable Diffusion's general effects are evident. Despite its limitations, it can produce beautiful photographs in some settings. Readers who want to use Stable Diffusion results should be warned. Source code examination demonstrates that Stable Diffusion images feature a concealed watermark (text StableDiffusionV1 and SDV2) encoded using the invisible-watermark Python package. It's not a secret, because the official Stable Diffusion repository's test watermark.py file contains a decoding snippet. The put watermark line in the txt2img.py source code can be removed if desired. I didn't discover this watermark on photographs made by the online Hugging Face demo. Maybe I did something incorrectly (but maybe they are just not using the txt2img script on their backend at all).

Conclusion

The Stable Diffusion model was fascinating. As I mentioned before, trying something yourself is always better than taking someone else's word, so I encourage readers to do the same (including this article as well;).

Is Generative AI a game-changer? My humble experience tells me:

• I think that place has a lot of potential. For designers and artists, generative AI can be a truly useful and innovative tool. Unfortunately, it can also pose a threat to some of them since if users can enter a text field to obtain a picture or a website logo in a matter of clicks, why would they pay more to a different party? Is it possible right now? unquestionably not yet. Images still have a very poor quality and are erroneous in minute details. And after viewing the image of the stunning woman above, models and fashion photographers may also unwind because it is highly unlikely that AI will replace them in the upcoming years.

• Today, generative AI is still in its infancy. Even 768x768 images are considered to be of a high resolution when using neural networks, which are computationally highly expensive. There isn't an AI model that can generate high-resolution photographs natively without upscaling or other methods, at least not as of the time this article was written, but it will happen eventually.

• It is still a challenge to accurately represent knowledge in neural networks (information like how many legs a cat has or the year Napoleon was born). Consequently, AI models struggle to create photorealistic photos, at least where little details are important (on the other side, when I searched Google for modern art paintings, the results are often even worse;).

• When compared to the carefully chosen images from official web pages or YouTube reviews, the average output quality of a Stable Diffusion generation process is actually less attractive because to its high degree of randomness. When using the same technique on their own, consumers will theoretically only view those images as 1% of the results.

Anyway, it's exciting to witness this area's advancement, especially because the project is open source. Google's Imagen and DALL-E 2 can also produce remarkable findings. It will be interesting to see how they progress.