Reflections on the closed company Hola! Dating app

I'll discuss my fuckups as an entrepreneur and CEO. All of them refer to the dating app Hola!, which I co-founded and starred in.

Spring 2021 was when we started. Two techies and two non-techies created a dating app. Pokemon Go and Tinder were combined.

Online dating is a business, and it takes two weeks from a like to a date. We questioned online dating app users if they met anyone offline last year.

75% replied yes, 50% sometimes, 25% usually.

Offline dating is popular, yet people have concerns.

• Men are reluctant to make mistakes in front of others.

• Women are curious about the background of everyone who approaches them.

We designed unique mechanics that let people date after a match. No endless chitchat. Women would be safe while men felt like cowboys.

I wish to emphasize three faults that lead to founders' estrangement.

This detachment ultimately led to us shutting down the company.

The wrong technology stack

Situation

Instead of generating a faster MVP and designing an app in a universal stack for iOS and Android, I argued we should pilot the app separately for iOS and Android. Technical founders' expertise made this possible.

Self-reflection

Mistaken strategy. We lost time and resources developing two apps at once. We chose iOS since it's more profitable. Apple took us out after the release, citing Guideline 4.3 Spam. After 4 months, we had nothing. We had a long way to go to get the app on Android and the Store.

I suggested creating a uniform platform for the company's growth. This makes parallel product development easier. The strategist's lack of experience and knowledge made it a piece of crap.

What would I have changed if I could?

We should have designed an Android universal stack. I expected Apple to have issues with a dating app.

Our approach should have been to launch something and subsequently improve it, but prejudice won.

The lesson

Discuss the IT stack with your CTO. It saves time and money. Choose the easiest MVP method.

2. A tardy search for investments

Situation

Though the universe and other founders encouraged me to locate investors first, I started pitching when we almost had an app.

When angels arrived, it was time to close. The app was banned, war broke out, I left the country, and the other co-founders stayed. We had no savings.

Self-reflection

I loved interviewing users. I'm proud of having done 1,000 interviews. I wanted to understand people's pain points and improve the product.

Interview results no longer affected the product. I was terrified to start pitching. I filled out accelerator applications and redid my presentation. You must go through that so you won't be terrified later.

What would I have changed if I could?

Get an external or internal mentor to help me with my first pitch as soon as possible. I'd be supported if criticized. He'd cheer with me if there was enthusiasm.

In 99% of cases, I'm comfortable jumping into the unknown, but there are exceptions. The mentor's encouragement would have prompted me to act sooner.

The lesson

Begin fundraising immediately. Months may pass. Show investors your pre-MVP project. Draw inferences from feedback.

3. Role ambiguity

Situation

My technical co-founders were also part-time lead developers, which produced communication issues. As co-founders, we communicated well and recognized the problems. Stakes, vesting, target markets, and approach were agreed upon.

We were behind schedule. Technical debt and strategic gap grew.

Bi-daily and weekly reviews didn't help. Each time, there were explanations. Inside, I was freaking out.

Self-reflection

I am a fairly easy person to talk to. I always try to stick to agreements; otherwise, my head gets stuffed with unnecessary information, interpretations, and emotions.

Sit down -> talk -> decide -> do -> evaluate the results. Repeat it.

If I don't get detailed comments, I start ruining everyone's mood. If there's a systematic violation of agreements without a good justification, I won't join the project or I'll end the collaboration.

What would I have done otherwise?

This is where it’s scariest to draw conclusions. Probably the most logical thing would have been not to start the project as we started it. But that was already a completely different project. So I would not have done anything differently and would have failed again.

But I drew conclusions for the future.

The lesson

First-time founders should find an adviser or team coach for a strategic session. It helps split the roles and responsibilities.

Most don't understand the issue until it's too late.

My first startup was what entrepreneurs call the hardest. A two-sided marketplace.

Two-sided marketplaces are the hardest startups because founders must solve the chicken or the egg conundrum.

A two-sided marketplace needs suppliers and buyers. Without suppliers, buyers won't come. Without buyers, suppliers won't come. An empty marketplace and a founder striving to gain momentum result.

My first venture made me a struggling founder seeking to achieve traction for a two-sided marketplace. The company failed, and I vowed never to start another like it.

I didn’t. Unfortunately, my second venture was almost as hard. It failed like the second-hardest startup.

What kind of startup is the second-hardest?

The second-hardest startup, which is almost as hard to develop, is rarely discussed in the startup community. Because of this, I predict more founders fail each year trying to develop the second-toughest startup than the hardest.

Fairly, I have no proof. I see many startups, so I have enough of firsthand experience. From what I've seen, for every entrepreneur developing a two-sided marketplace, I'll meet at least 10 building this other challenging startup.

I'll describe a startup I just met with its two co-founders to explain the second hardest sort of startup and why it's so hard. They created a financial literacy software for parents of high schoolers.

The issue appears plausible. Children struggle with money. Parents must teach financial responsibility. Problems?

It's possible.

Buyers and users are different.

Buyer-user mismatch.

The financial literacy app I described above targets parents. The parent doesn't utilize the app. Child is end-user. That may not seem like much, but it makes customer and user acquisition and onboarding difficult for founders.

The difficulty of a buyer-user imbalance

The company developing a product faces a substantial operational burden when the buyer and end customer are different. Consider classic firms where the buyer is the end user to appreciate that responsibility.

Entrepreneurs selling directly to end users must educate them about the product's benefits and use. Each demands a lot of time, effort, and resources.

Imagine selling a financial literacy app where the buyer and user are different. To make the first sale, the entrepreneur must establish all the items I mentioned above. After selling, the entrepreneur must supply a fresh set of resources to teach, educate, or train end-users.

Thus, a startup with a buyer-user mismatch must market, sell, and train two organizations at once, requiring twice the work with the same resources.

The second hardest startup is hard for reasons other than the chicken-or-the-egg conundrum. It takes a lot of creativity and luck to solve the chicken-or-egg conundrum.

The buyer-user mismatch problem cannot be overcome by innovation or luck. Buyer-user mismatches must be solved by force. Simply said, when a product buyer is different from an end-user, founders have a lot more work. If they can't work extra, their companies fail.

Sundar Pichai unveiled Simplicity Sprint at Google's latest all-hands conference.

To boost employee efficiency.

Not surprising. Few envisioned Google declaring a productivity drive.

Sunder Pichai's speech:

“There are real concerns that our productivity as a whole is not where it needs to be for the head count we have. Help me create a culture that is more mission-focused, more focused on our products, more customer focused. We should think about how we can minimize distractions and really raise the bar on both product excellence and productivity.”

The primary driver driving Google's efficiency push is:

Google's efficiency push follows 13% quarterly revenue increase. Last year in the same quarter, it was 62%.

Market newcomers may argue that the previous year's figure was fuelled by post-Covid reopening and growing consumer spending. Investors aren't convinced. A promising company like Google can't afford to drop so quickly.

Google’s quarterly revenue growth stood at 13%, against 62% in last year same quarter.

Google isn't alone. In my recent essay regarding 2025 programmers, I warned about the economic downturn's effects on FAAMG's workforce. Facebook had suspended hiring, and Microsoft had promised hefty bonuses for loyal staff.

In the same article, I predicted Google's troubles. Online advertising, especially the way Google and Facebook sell it using user data, is over.

FAAMG and 2nd rung IT companies could be the first to fall without Post-COVID revival and uncertain global geopolitics.

Google has hardly ever discussed effectiveness:

Apparently openly.

Amazon treats its employees like robots, even in software positions. It has significant turnover and a terrible reputation as a result. Because of this, it rarely loses money due to staff productivity.

Amazon trumps Google. In reality, it treats its employees poorly.

Google was the founding father of the modern-day open culture.

Larry and Sergey Google founded the IT industry's Open Culture. Silicon Valley called Google's internal democracy and transparency near anarchy. Management rarely slammed decisions on employees. Surveys and internal polls ensured everyone knew the company's direction and had a vote.

20% project allotment (weekly free time to build own project) was Google's open-secret innovation component.

After Larry and Sergey's exit in 2019, this is Google's first profitability hurdle. Only Google insiders can answer these questions.

• Would Google's investors compel the company's management to adopt an Amazon-style culture where the developers are treated like circus performers?

• If so, would Google follow suit?

• If so, how does Google go about doing it?

Before discussing Google's likely plan, let's examine programming productivity.

What determines a programmer's productivity is simple:

How would we answer Google's questions?

As a programmer, I'm more concerned about Simplicity Sprint's aftermath than its economic catalysts.

Large organizations don't care much about quarterly and annual productivity metrics. They have 10-year product-launch plans. If something seems horrible today, it's likely due to someone's lousy judgment 5 years ago who is no longer in the blame game.

Deconstruct our main question.

• How exactly do you change the culture of the firm so that productivity increases?

• How can you accomplish that without affecting your capacity to profit? There are countless ways to increase output without decreasing profit.

• How can you accomplish this with little to no effect on employee motivation? (While not all employers care about it, in this case we are discussing the father of the open company culture.)

• How do you do it for a 10-developer IT firm that is losing money versus a 1,70,000-developer organization with a trillion-dollar valuation?

When implementing a large-scale organizational change, success must be carefully measured.

The fastest way to do something is to do it right, no matter how long it takes.

You require clearly-defined group/team/role segregation and solid pass/fail matrices to:

• You can give performers rewards.

• Ones that are average can be inspired to improve

• Underachievers may receive assistance or, in the worst-case scenario, rehabilitation

As a 20-year programmer, I associate productivity with greatness.

Doing something well, no matter how long it takes, is the fastest way to do it.

Let's discuss a programmer's productivity.

Why productivity is a strange term in programming:

Productivity is work per unit of time.

Money=time This is an economic proverb. More hours worked, more pay. Longer projects cost more.

As a buyer, you desire a quick supply. As a business owner, you want employees who perform at full capacity, creating more products to transport and boosting your profits.

All economic matrices encourage production because of our obsession with it. Productivity is the only organic way a nation may increase its GDP.

Time is money — is not just a proverb, but an economical fact.

Applying the same productivity theory to programming gets problematic. An automating computer. Its capacity depends on the software its master writes.

Today, a sophisticated program can process a billion records in a few hours. Creating one takes a competent coder and the necessary infrastructure. Learning, designing, coding, testing, and iterations take time.

Programming productivity isn't linear, unlike manufacturing and maintenance.

Average programmers produce code every day yet miss deadlines. Expert programmers go days without coding. End of sprint, they often surprise themselves by delivering fully working solutions.

Reversing the programming duties has no effect. Experts aren't needed for productivity.

These patterns remind me of an XKCD comic.

Programming productivity depends on two factors:

• The capacity of the programmer and his or her command of the principles of computer science

• His or her productive bursts, how often they occur, and how long they last as they engineer the answer

At some point, productivity measurement becomes Schrödinger’s cat.

Product companies measure productivity using use cases, classes, functions, or LOCs (lines of code). In days of data-rich source control systems, programmers' merge requests and/or commits are the most preferred yardstick. Companies assess productivity by tickets closed.

Every organization eventually has trouble measuring productivity. Finer measurements create more chaos. Every measure compares apples to oranges (or worse, apples with aircraft.) On top of the measuring overhead, the endeavor causes tremendous and unnecessary stress on teams, lowering their productivity and defeating its purpose.

Macro productivity measurements make sense. Amazon's factory-era management has done it, but at great cost.

Google can pull it off if it wants to.

What Google meant in reality when it said that employee productivity has decreased:

When Google considers its employees unproductive, it doesn't mean they don't complete enough work in the allotted period.

They can't multiply their work's influence over time.

• Programmers who produce excellent modules or products are unsure on how to use them.

• The best data scientists are unable to add the proper parameters in their models.

• Despite having a great product backlog, managers struggle to recruit resources with the necessary skills.

• Product designers who frequently develop and A/B test newer designs are unaware of why measures are inaccurate or whether they have already reached the saturation point.

• Most ignorant: All of the aforementioned positions are aware of what to do with their deliverables, but neither their supervisors nor Google itself have given them sufficient authority.

So, Google employees aren't productive.

How to fix it?

• Business analysis: White suits introducing novel items can interact with customers from all regions. Track analytics events proactively, especially the infrequent ones.

• SOLID, DRY, TEST, and AUTOMATION: Do less + reuse. Use boilerplate code creation. If something already exists, don't implement it yourself.

• Build features-building capabilities: N features are created by average programmers in N hours. An endless number of features can be built by average programmers thanks to the fact that expert programmers can produce 1 capability in N hours.

• Work on projects that will have a positive impact: Use the same algorithm to search for images on YouTube rather than the Mars surface.

• Avoid tasks that can only be measured in terms of time linearity at all costs (if a task can be completed in N minutes, then M copies of the same task would cost M*N minutes).

In conclusion:

Software development isn't linear. Why should the makers be measured?

Notation for The Big O

I'm discussing a new way to quantify programmer productivity. (It applies to other professions, but that's another subject)

The Big O notation expresses the paradigm (the algorithmic performance concept programmers rot to ace their Google interview)

Google (or any large corporation) can do this.

1. Sort organizational roles into categories and specify their impact vs. time objectives. A CXO role's time vs. effect function, for instance, has a complexity of O(log N), meaning that if a CEO raises his or her work time by 8x, the result only increases by 3x.

2. Plot the influence of each employee over time using the X and Y axes, respectively.

3. Add a multiplier for Y-axis values to the productivity equation to make business objectives matter. (Example values: Support = 5, Utility = 7, and Innovation = 10).

4. Compare employee scores in comparable categories (developers vs. devs, CXOs vs. CXOs, etc.) and reward or help employees based on whether they are ahead of or behind the pack.

After measuring every employee's inventiveness, it's straightforward to help underachievers and praise achievers.

Example of a Big(O) Category:

If I ran Google (God forbid, its worst days are far off), here's how I'd classify it. You can categorize Google employees whichever you choose.

The Google interview truth:

O(1) < O(log n) < O(n) < O(n log n) < O(n^x) where all logarithmic bases are < n.

O(1): Customer service workers' hours have no impact on firm profitability or customer pleasure.

CXOs Most of their time is spent on travel, strategic meetings, parties, and/or meetings with minimal floor-level influence. They're good at launching new products but bad at pivoting without disaster. Their directions are being followed.

Devops, UX designers, testers Agile projects revolve around deployment. DevOps controls the levers. Their automation secures results in subsequent cycles.

UX/UI Designers must still prototype UI elements despite improved design tools.

All test cases are proportional to use cases/functional units, hence testers' work is O(N).

Architects Their effort improves code quality. Their right/wrong interference affects product quality and rollout decisions even after the design is set.

Core Developers Only core developers can write code and own requirements. When people understand and own their labor, the output improves dramatically. A single character error can spread undetected throughout the SDLC and cost millions.

Core devs introduce/eliminate 1000x bugs, refactoring attempts, and regression. Following our earlier hypothesis.

The fastest way to do something is to do it right, no matter how long it takes.

Conclusion:

Google is at the liberal extreme of the employee-handling spectrum

Microsoft faced an existential crisis after 2000. It didn't choose Amazon's data-driven people management to revitalize itself.

Instead, it entrusted developers. It welcomed emerging technologies and opened up to open source, something it previously opposed.

Google is too lax in its employee-handling practices. With that foundation, it can only follow Amazon, no matter how carefully.

Any attempt to redefine people's measurements will affect the organization emotionally.

The more Google compares apples to apples, the higher its chances for future rebirth.

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.