Many developers do it, but I don't approve.

Aren't many programmers part-time? Many work full-time but also freelance. If the job agreement allows it, I see no problem.

Tech businesses' policies vary. I have a friend in Google, Germany. According to his contract, he couldn't do an outside job. Google owns any code he writes while employed.

I was shocked. Later, I found that different Google regions have different policies.

A corporation can normally establish any agreement before hiring you. They're negotiable. When there's no agreement, state law may apply. In court, law isn't so simple.

I won't delve into legal details. Instead, let’s talk about the incident.

How he was discovered

In one month, he missed two deadlines. His boss was frustrated because the assignment wasn't difficult to miss twice. When a team can't finish work on time, they all earn bad grades.

He annoyed the whole team. One team member (anonymous) told the project manager he worked on side projects during office hours. He may have missed deadlines because of this.

The project manager was furious. He needed evidence. The manager caught him within a week. The manager told higher-ups immediately.

The company wanted to set an example

Management could terminate him and settle the problem. But the company wanted to set an example for those developers who breached the regulation.

Because dismissal isn't enough. Every organization invests heavily in developer hiring. If developers depart or are fired after a few months, the company suffers.

The developer spent 10 months there. The employer sacked him and demanded ten months' pay. Or they'd sue him.

It was illegal and unethical. The youngster paid the fine and left the company quietly to protect his career.

Right or wrong?

Is the developer's behavior acceptable? Let's discuss developer malpractice.

During office hours, may developers work on other projects? If they're bored during office hours, they might not. Check the employment contract or state law.

If there's no employment clause, check country/state law. Because you can't justify breaking the law. Always. Most employers own their employees' work hours unless it's a contractual position.

If the company agrees, it's fine.

I also oppose companies that force developers to work overtime without pay.

Most states and countries have laws that help companies and workers. Law supports employers in this case. If any of the following are true, the company/employer owns the IP under California law.

• using the business's resources

• any equipment, including a laptop used for business.

• company's mobile device.

• offices of the company.

• business time as well. This is crucial. Because this occurred in the instance of my junior.

Company resources are dangerous. Because your company may own the product's IP.  If you have seen the TV show Silicon Valley, you have seen a similar situation there, right?

Conclusion

Simple rule. I avoid big side projects. I work on my laptop on weekends for side projects. I'm safe. But I also know that my company might not be happy with that.

As an employee, I suppose I can. I can make side money. I won't promote it, but I'll respect their time, resources, and task. I also sometimes work extra time to finish my company’s deadlines.

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.

There’s a funny part of ‘Steve Jobs’ by Walter Isaacson where Jobs says that Bill Gates was more a copier than an innovator:

“Bill is basically unimaginative and has never invented anything, which is why I think he’s more comfortable now in philanthropy than technology. He just shamelessly ripped off other people’s ideas….He’d be a broader guy if he had dropped acid once or gone off to an ashram when he was younger.”

Gates lacked flavor. Nobody ever got excited about a Microsoft launch, despite their good products. Jobs had the world's best product taste. Apple vs. Microsoft.

A CEO's core job functions are all driven by taste: recruiting, vision, and company culture all require good taste. Depending on the type of company you want to build, know where you stand between Microsoft and Apple.

How can you improve your product judgment? How to acquire taste?

Test and refine

Product development follows two parallel paths: the ‘customer obsession’ path and the ‘taste and iterate’ path.

The customer obsession path involves solving customer problems. Lean Startup frameworks show you what to build at each step.

Taste-and-iterate doesn't involve the customer. You iterate internally and rely on product leaders' taste and judgment.

Creative Selection by Ken Kocienda explains this method. In Creative Selection, demos are iterated and presented to product leaders. Your boss presents to their boss, and so on up to Steve Jobs. If you have good product taste, you can be a panelist.

The iPhone follows this path. Before seeing an iPhone, consumers couldn't want one. Customer obsession wouldn't have gotten you far because iPhone buyers didn't know they wanted one.

In The Hard Thing About Hard Things, Ben Horowitz writes:

“It turns out that is exactly what product strategy is all about — figuring out the right product is the innovator’s job, not the customer’s job. The customer only knows what she thinks she wants based on her experience with the current product. The innovator can take into account everything that’s possible, but often must go against what she knows to be true. As a result, innovation requires a combination of knowledge, skill, and courage.“

One path solves a problem the customer knows they have, and the other doesn't. Instead of asking a person what they want, observe them and give them something they didn't know they needed.

It's much harder. Apple is the world's most valuable company because it's more valuable. It changes industries permanently.

If you want to build superior products, use the iPhone of your industry.

How to Improve Your Taste

I. Work for a company that has taste.

People with the best taste in products, markets, and people are rewarded for building great companies. Tasteful people know quality even when they can't describe it. Taste isn't writable. It's feel-based.

Moving into a community that's already doing what you want to do may be the best way to develop entrepreneurial taste. Most company-building knowledge is tacit.

Joining a company you want to emulate allows you to learn its inner workings. It reveals internal patterns intuitively. Many successful founders come from successful companies.

Consumption determines taste. Excellence will refine you. This is why restauranteurs visit the world's best restaurants and serious painters visit Paris or New York. Joining a company with good taste is beneficial.

2. Possess a wide range of interests

“Edwin Land of Polaroid talked about the intersection of the humanities and science. I like that intersection. There’s something magical about that place… The reason Apple resonates with people is that there’s a deep current of humanity in our innovation. I think great artists and great engineers are similar, in that they both have a desire to express themselves.” — Steve Jobs

I recently discovered Edwin Land. Jobs modeled much of his career after Land's. It makes sense that Apple was inspired by Land.

A Triumph of Genius: Edwin Land, Polaroid, and the Kodak Patent War notes:

“Land was introverted in person, but supremely confident when he came to his ideas… Alongside his scientific passions, lay knowledge of art, music, and literature. He was a cultured person growing even more so as he got older, and his interests filtered into the ethos of Polaroid.”

Founders' philosophies shape companies. Jobs and Land were invested. It showed in the products their companies made. Different. His obsession was spreading Microsoft software worldwide. Microsoft's success is why their products are bland and boring.

Experience is important. It's probably why startups are built by generalists and scaled by specialists.

Jobs combined design, typography, storytelling, and product taste at Apple. Some of the best original Mac developers were poets and musicians. Edwin Land liked broad-minded people, according to his biography. Physicist-musicians or physicist-photographers.

Da Vinci was a master of art, engineering, architecture, anatomy, and more. He wrote and drew at the same desk. His genius is remembered centuries after his death. Da Vinci's statue would stand at the intersection of humanities and science.

We find incredibly creative people here. Superhumans. Designers, creators, and world-improvers. These are the people we need to navigate technology and lead world-changing companies. Generalists lead.

You have to use it.

Chat-GPT is a revolution.

All social media outlets are discussing it. How it will impact the future and different things.

True.

I've been using Chat-GPT for a few days, and it's a rare revolution. It's amazing and will only improve.

I asked Chat-GPT about the best non-fiction books. It advised this, albeit results rely on interests.

The Immortal Life of Henrietta Lacks

by Rebecca Skloot

Science, Biography

A impoverished tobacco farmer dies of cervical cancer in The Immortal Life of Henrietta Lacks. Her cell strand helped scientists treat polio and other ailments.

Rebecca Skloot discovers about Henrietta, her family, how the medical business exploited black Americans, and how her cells can live forever in a fascinating and surprising research.

You ought to read it.

1. if you want to discover more about the past of medicine.

2. if you want to discover more about American history.

Bad Blood: Secrets and Lies in a Silicon Valley Startup

by John Carreyrou

Tech, Bio

Bad Blood tells the terrifying story of how a Silicon Valley tech startup's blood-testing device placed millions of lives at risk.

John Carreyrou, a Pulitzer Prize-winning journalist, wrote this book.

Theranos and its wunderkind CEO, Elizabeth Holmes, climbed to popularity swiftly and then plummeted.

You ought to read it.

1. if you are a start-up employee.

2. specialists in medicine.

The Power of Now: A Guide to Spiritual Enlightenment

by Eckhart Tolle

Self-improvement, Spirituality

The Power of Now shows how to stop suffering and attain inner peace by focusing on the now and ignoring your mind.

The book also helps you get rid of your ego, which tries to control your ideas and actions.

If you do this, you may embrace the present, reduce discomfort, strengthen relationships, and live a better life.

You ought to read it.

1. if you're looking for serenity and illumination.

2. If you believe that you are ruining your life, stop.

3. if you're not happy.

The 7 Habits of Highly Effective People

by Stephen R. Covey

Profession, Success

The 7 Habits of Highly Effective People is an iconic self-help book.

This vital book offers practical guidance for personal and professional success.

This non-fiction book is one of the most popular ever.

You ought to read it.

1. if you want to reach your full potential.

2. if you want to discover how to achieve all your objectives.

3. if you are just beginning your journey toward personal improvement.

Sapiens: A Brief History of Humankind

by Yuval Noah Harari

Science, History

Sapiens explains how our species has evolved from our earliest ancestors to the technology age.

How did we, a species of hairless apes without tails, come to control the whole planet?

It describes the shifts that propelled Homo sapiens to the top.

You ought to read it.

1. if you're interested in discovering our species' past.

2. if you want to discover more about the origins of human society and culture.

The 2008 Bitcoin white paper solves the classic computer science consensus problem.

Issue Statement

The Byzantine Generals Problem (BGP) is called after an allegory in which several generals must collaborate and attack a city at the same time to win (figure 1-left). Any general who retreats at the last minute loses the fight (figure 1-right). Thus, precise messengers and no rogue generals are essential. This is difficult without a trusted central authority.

In their 1982 publication, Leslie Lamport, Robert Shostak, and Marshall Please termed this topic the Byzantine Generals Problem to simplify distributed computer systems.

Consensus in a distributed computer network is the issue. Reaching a consensus on which systems work (and stay in the network) and which don't makes maintaining a network tough (i.e., needs to be removed from network). Challenges include unreliable communication routes between systems and mis-reporting systems.

Solving BGP can let us construct machine learning solutions without single points of failure or trusted central entities. One server hosts model parameters while numerous workers train the model. This study describes fault-tolerant Distributed Byzantine Machine Learning.

Bitcoin invented a mechanism for a distributed network of nodes to agree on which transactions should go into the distributed ledger (blockchain) without a trusted central body. It solved BGP implementation. Satoshi Nakamoto, the pseudonymous bitcoin creator, solved the challenge by cleverly combining cryptography and consensus mechanisms.

Disclaimer

This is not financial advice. It discusses a unique computer science solution.

Bitcoin

Bitcoin's white paper begins:

“A purely peer-to-peer version of electronic cash would allow online payments to be sent directly from one party to another without going through a financial institution.” Source: https://www.ussc.gov/sites/default/files/pdf/training/annual-national-training-seminar/2018/Emerging_Tech_Bitcoin_Crypto.pdf

Bitcoin's main parts:

1. The open-source and versioned bitcoin software that governs how nodes, miners, and the bitcoin token operate.

2. The native kind of token, known as a bitcoin token, may be created by mining (up to 21 million can be created), and it can be transferred between wallet addresses in the bitcoin network.

3. Distributed Ledger, which contains exact copies of the database (or "blockchain") containing each transaction since the first one in January 2009.

4. distributed network of nodes (computers) running the distributed ledger replica together with the bitcoin software. They broadcast the transactions to other peer nodes after validating and accepting them.

5. Proof of work (PoW) is a cryptographic requirement that must be met in order for a miner to be granted permission to add a new block of transactions to the blockchain of the cryptocurrency bitcoin. It takes the form of a valid hash digest. In order to produce new blocks on average every 10 minutes, Bitcoin features a built-in difficulty adjustment function that modifies the valid hash requirement (length of nonce). PoW requires a lot of energy since it must continually generate new hashes at random until it satisfies the criteria.

6. The competing parties known as miners carry out continuous computing processing to address recurrent cryptography issues. Transaction fees and some freshly minted (mined) bitcoin are the rewards they receive. The amount of hashes produced each second—or hash rate—is a measure of mining capacity.

Cryptography, decentralization, and the proof-of-work consensus method are Bitcoin's most unique features.

Bitcoin uses encryption

Bitcoin employs this established cryptography.

1. Hashing

2. digital signatures based on asymmetric encryption

Hashing (SHA-256) (SHA-256)

Hashing converts unique plaintext data into a digest. Creating the plaintext from the digest is impossible. Bitcoin miners generate new hashes using SHA-256 to win block rewards.

A new hash is created from the current block header and a variable value called nonce. To achieve the required hash, mining involves altering the nonce and re-hashing.

The block header contains the previous block hash and a Merkle root, which contains hashes of all transactions in the block. Thus, a chain of blocks with increasing hashes links back to the first block. Hashing protects new transactions and makes the bitcoin blockchain immutable. After a transaction block is mined, it becomes hard to fabricate even a little entry.

Asymmetric Cryptography Digital Signatures

Asymmetric cryptography (public-key encryption) requires each side to have a secret and public key. Public keys (wallet addresses) can be shared with the transaction party, but private keys should not. A message (e.g., bitcoin payment record) can only be signed by the owner (sender) with the private key, but any node or anybody with access to the public key (visible in the blockchain) can verify it. Alex will submit a digitally signed transaction with a desired amount of bitcoin addressed to Bob's wallet to a node to send bitcoin to Bob. Alex alone has the secret keys to authorize that amount. Alex's blockchain public key allows anyone to verify the transaction.

Solution

Now, apply bitcoin to BGP. BGP generals resemble bitcoin nodes. The generals' consensus is like bitcoin nodes' blockchain block selection. Bitcoin software on all nodes can:

Check transactions (i.e., validate digital signatures)

2. Accept and propagate just the first miner to receive the valid hash and verify it accomplished the task. The only way to guess the proper hash is to brute force it by repeatedly producing one with the fixed/current block header and a fresh nonce value.

Thus, PoW and a dispersed network of nodes that accept blocks from miners that solve the unfalsifiable cryptographic challenge solve consensus.

Suppose:

1. Unreliable nodes

2. Unreliable miners

Bitcoin accepts the longest chain if rogue nodes cause divergence in accepted blocks. Thus, rogue nodes must outnumber honest nodes in accepting/forming the longer chain for invalid transactions to reach the blockchain. As of November 2022, 7000 coordinated rogue nodes are needed to takeover the bitcoin network.

Dishonest miners could also try to insert blocks with falsified transactions (double spend, reverse, censor, etc.) into the chain. This requires over 50% (51% attack) of miners (total computational power) to outguess the hash and attack the network. Mining hash rate exceeds 200 million (source). Rewards and transaction fees encourage miners to cooperate rather than attack. Quantum computers may become a threat.

Visit my Quantum Computing post.

Quantum computers—what are they? Quantum computers will have a big influence. towardsdatascience.com

Nodes have more power than miners since they can validate transactions and reject fake blocks. Thus, the network is secure if honest nodes are the majority.

Summary

Table 1 compares three Byzantine Generals Problem implementations.

Bitcoin white paper and implementation solved the consensus challenge of distributed systems without central governance. It solved the illusive Byzantine Generals Problem.

Resources

Resources

1. https://en.wikipedia.org/wiki/Byzantine_fault

2. Source-code for Bitcoin Core Software — https://github.com/bitcoin/bitcoin

3. Bitcoin white paper — https://bitcoin.org/bitcoin.pdf

4. https://en.wikipedia.org/wiki/Bitcoin

5. https://www.microsoft.com/en-us/research/publication/byzantine-generals-problem/

6. https://www.microsoft.com/en-us/research/uploads/prod/2016/12/The-Byzantine-Generals-Problem.pdf

7. https://en.wikipedia.org/wiki/Hash_function

8. https://en.wikipedia.org/wiki/Merkle_tree

9. https://en.wikipedia.org/wiki/SHA-2

10. https://en.wikipedia.org/wiki/Public-key_cryptography

11. https://en.wikipedia.org/wiki/Digital_signature

12. https://en.wikipedia.org/wiki/Proof_of_work

13. https://en.wikipedia.org/wiki/Quantum_cryptography

14. https://dci.mit.edu/bitcoin-security-initiative

15. https://dci.mit.edu/51-attacks

16. Genuinely Distributed Byzantine Machine Learning, El-Mahdi El-Mhamdi et al., 2020. ACM, New York, NY, https://doi.org/10.1145/3382734.3405695

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.