My growth-killing mistake

When you initially start developing your audience, you need guidance.

What does my audience like? What do they not like? How can I grow more?

When I started writing two years ago, I inquired daily. Taking cues from your audience to develop more valuable content is a good concept, but it's simple to let them destroy your growth.

A small audience doesn't represent the full picture.

When I had fewer than 100 YouTube subscribers, I tried several video styles and topics. I looked to my audience for what to preserve and what to change.

If my views, click-through rate, or average view % dropped, that topic or style was awful. Avoiding that style helped me grow.

Vlogs, talking head videos on writing, and long-form tutorials didn't fare well.

Since I was small, I've limited the types of films I make. I have decided to make my own videos.

Surprisingly, the videos I avoided making meet or exceed my views, CTR, and audience retention.

A limited audience can't tell you what your tribe wants. Therefore, limiting your innovation will prohibit you from reaching the right audience. Finding them may take longer.

Large Creators Experience The Same Issue

In the last two years, I've heard Vanessa Lau and Cathrin Manning say they felt pigeonholed into generating videos they didn't want to do.

Why does this happen over and over again?

Once you have a popular piece of content, your audience will grow. So when you publish inconsistent material, fewer of your new audience will view it. You interpret the drop in views as a sign that your audience doesn't want the content, so you stop making it.

Repeat this procedure a few times, and you'll create stuff you're not passionate about because you're frightened to publish it.

How to Manage Your Creativity and Audience Development

I'm not recommending you generate random content.

Instead of feeling trapped by your audience, you can cultivate a diverse audience.

Create quality material on a range of topics and styles as you improve. Be creative until you get 100 followers. Look for comments on how to improve your article.

If you observe trends in the types of content that expand your audience, focus 50-75% of your material on those trends. Allow yourself to develop 25% non-performing material.

This method can help you expand your audience faster with your primary trends and like all your stuff. Slowly, people will find 25% of your material, which will boost its performance.

How to Expand Your Audience Without Having More Limited Content

Follow these techniques to build your audience without feeling confined.

• Don't think that you need restrict yourself to what your limited audience prefers.

• Don't let the poor performance of your desired material demotivate you.

• You shouldn't restrict the type of content you publish or the themes you cover when you have less than 100 followers.

• When your audience expands, save 25% of your content for your personal interests, regardless of how well it does.

One-hour ebook sales, affiliate relationships, and more.

If you follow me, you may know I started a new ebook in August 2022.

Despite publishing on this platform, my website, and Quora, I'm not a writer.

My writing speed is slow, 2,000 words a day, and I struggle to communicate cohesively.

In April 2022, I wrote a successful guide on How to Write Google-Friendly Blog Posts.

I had no email list or social media presence. I've made $1,600+ selling ebooks.

Evidence:

My first digital offering isn't a book.

It's an actionable guide with my tried-and-true process for writing Google-friendly content.

I'm not bragging.

Established authors like Tim Denning make more from my ebook sales with one newsletter.

This experience taught me writing isn't a privilege.

Writing a book and making money online doesn't require expertise.

Many don't consult experts. They want someone approachable.

Two years passed before I realized my own limits.

I have a brain, two hands, and Internet to spread my message.

I wrote and published a second ebook after the first's success.

On Gumroad, I released my second digital product.

Here's my complete Gumroad evaluation.

Gumroad is a marketplace for content providers to develop and sell sales pages.

Gumroad handles payments and client requests. It's helpful when someone sends a bogus payment receipt requesting an ebook (actual story!).

You'll forget administrative concerns after your first ebook sale.

After my first ebook sale, I did this: I made additional cash!

After every sale, I tell myself, "I built a new semi-passive revenue source."

This thinking shift helps me become less busy while increasing my income and quality of life.

Besides helping others, folks sell evergreen digital things to earn passive money.

It's in my second ebook.

I explain how I built and sold 50+ copies of my SEO writing ebook without being an influencer.

I show how anyone can sell ebooks on Gumroad and automate their sales process.

This is my ebook.

After publicizing the ebook release, I sold three copies within an hour.

Wow, or meh?

I don’t know.

The answer is different for everyone.

These three sales came from a small email list of 40 motivated fans waiting for my ebook release.

I had bigger plans.

I'll market my ebook on Medium, my website, Quora, and email.

I'm testing affiliate partnerships this time.

One of my ebook buyers is now promoting it for 40% commission.

Become my affiliate if you think your readers would like my ebook.

My ebook is a few days old, but I'm interested to see where it goes.

My SEO writing book started without an email list, affiliates, or 4,000 website visitors. I've made four figures.

I'm slowly expanding my communication avenues to have more impact.

Even a small project can open doors you never knew existed.

So began my writing career.

In summary

If you dare, every concept can become a profitable trip.

Before, I couldn't conceive of creating an ebook.

How to Sell eBooks on Gumroad is my second digital product.

Marketing and writing taught me that anything can be sold online.

Open to ideas or acquisitions

Failure is an unavoidable aspect of life, yet many recoil at the word.

I've worked on unrelated startup projects. This is a list of products I developed (often as the tech lead or co-founder) and why they failed to launch.

Chess Bet (Betting)

As a chess player who plays 5 games a day and has an ELO rating of 2100, I tried to design a chess engine to rival stockfish and Houdini.

While constructing my chess engine, my cofounder asked me about building a p2p chess betting app. Chess Bet. There couldn't be a better time.

Two people in different locations could play a staked game. The winner got 90% of the bet and we got 10%. The business strategy was clear, but our mini-launch was unusual.

People started employing the same cheat engines I mentioned, causing user churn and defaming our product.

It was the first programming problem I couldn't solve after building a cheat detection system based on player move strengths and prior games. Chess.com, the most famous online chess software, still suffers from this.

We decided to pivot because we needed an expensive betting license.

We relaunched as Chess MVP after deciding to focus on chess learning. A platform for teachers to create chess puzzles and teach content. Several chess students used our product, but the target market was too tiny.

We chose to quit rather than persevere or pivot.

BodaCare (Insure Tech)

‘BodaBoda’ in Swahili means Motorcycle. My Dad approached me in 2019 (when I was working for a health tech business) about establishing an Insurtech/fintech solution for motorbike riders to pay for insurance using SNPL.

We teamed up with an underwriter to market motorcycle insurance. Once they had enough premiums, they'd get an insurance sticker in the mail. We made it better by splitting the cover in two, making it more reasonable for motorcyclists struggling with lump-sum premiums.

Lack of capital and changing customer behavior forced us to close, with 100 motorcyclists paying 0.5 USD every day. Our unit econ didn't make sense, and CAC and retention capital only dug us deeper.

Circle (Social Networking)

Having learned from both product failures, I began to understand what worked and what didn't. While reading through Instagram, an idea struck me.

Suppose social media weren't virtual.

Imagine meeting someone on your way home. Like-minded person

People were excited about social occasions after covid restrictions were eased. Anything to escape. I just built a university student-popular experiences startup. Again, there couldn't be a better time.

I started the Android app. I launched it on Google Beta and oh my! 200 people joined in two days.

It works by signaling if people are in a given place and allowing users to IM in hopes of meeting up in near real-time. Playstore couldn't deploy the app despite its success in beta for unknown reasons. I appealed unsuccessfully.

My infrastructure quickly lost users because I lacked funding.

In conclusion

This essay contains many failures, some of which might have been avoided and others not, but they were crucial learning points in my startup path.

If you liked any idea, I have the source code on Github.

Happy reading until then!

ERC165 (or EIP-165) is a standard utilized by various open-source smart contracts like Open Zeppelin or Aavegotchi.

What's it? You must implement? Why do we need it? I'll describe the standard and answer any queries.

What is ERC165

ERC165 detects and publishes smart contract interfaces. Meaning? It standardizes how interfaces are recognized, how to detect if they implement ERC165, and how a contract publishes the interfaces it implements. How does it work?

Why use ERC165? Sometimes it's useful to know which interfaces a contract implements, and which version.

Identifying interfaces

An interface function's selector. This verifies an ABI function. XORing all function selectors defines an interface in this standard. The following code demonstrates.

// SPDX-License-Identifier: UNLICENCED
pragma solidity >=0.8.0 <0.9.0;

interface Solidity101 {
    function hello() external pure;
    function world(int) external pure;
}

contract Selector {
    function calculateSelector() public pure returns (bytes4) {
        Solidity101 i;
        return i.hello.selector ^ i.world.selector;
        // Returns 0xc6be8b58
    }

    function getHelloSelector() public pure returns (bytes4) {
        Solidity101 i;
        return i.hello.selector;
        // Returns 0x19ff1d21
    }

    function getWorldSelector() public pure returns (bytes4) {
        Solidity101 i;
        return i.world.selector;
        // Returns 0xdf419679
    }
}

This code isn't necessary to understand function selectors and how an interface's selector can be determined from the functions it implements.

Run that sample in Remix to see how interface function modifications affect contract function output.

Contracts publish their implemented interfaces.

We can identify interfaces. Now we must disclose the interfaces we're implementing. First, import IERC165 like so.

pragma solidity ^0.4.20;

interface ERC165 {
    /// @notice Query if a contract implements an interface
    /// @param interfaceID The interface identifier, as specified in ERC-165
    /// @dev Interface identification is specified in ERC-165. 
    /// @return `true` if the contract implements `interfaceID` and
    ///  `interfaceID` is not 0xffffffff, `false` otherwise
    function supportsInterface(bytes4 interfaceID) external view returns (bool);
}

We still need to build this interface in our smart contract. ERC721 from OpenZeppelin is a good example.

// SPDX-License-Identifier: MIT
// OpenZeppelin Contracts (last updated v4.5.0) (token/ERC721/ERC721.sol)

pragma solidity ^0.8.0;

import "./IERC721.sol";
import "./extensions/IERC721Metadata.sol";
import "../../utils/introspection/ERC165.sol";
// ...

contract ERC721 is Context, ERC165, IERC721, IERC721Metadata {
  // ...

  function supportsInterface(bytes4 interfaceId) public view virtual override(ERC165, IERC165) returns (bool) {
    return
      interfaceId == type(IERC721).interfaceId ||
      interfaceId == type(IERC721Metadata).interfaceId ||
      super.supportsInterface(interfaceId);
  }
  
  // ...
}

I deleted unnecessary code. The smart contract imports ERC165, IERC721 and IERC721Metadata. The is keyword at smart contract declaration implements all three.

Kind (interface).

Note that type(interface).interfaceId returns the same as the interface selector.

We override supportsInterface in the smart contract to return a boolean that checks if interfaceId is the same as one of the implemented contracts.

Super.supportsInterface() calls ERC165 code. Checks if interfaceId is IERC165.

function supportsInterface(bytes4 interfaceId) public view virtual override returns (bool) {
    return interfaceId == type(IERC165).interfaceId;
}

So, if we run supportsInterface with an interfaceId, our contract function returns true if it's implemented and false otherwise. True for IERC721, IERC721Metadata, andIERC165.

Conclusion

I hope this post has helped you understand and use ERC165 and why it's employed.

Have a great day, thanks for reading!

No other calendar is simpler, smaller, and reusable year after year. It works and is used here.

Most of us discard and replace our calendars annually. Each month, we move our calendar ahead another page, thus if we need to know which day of the week corresponds to a given day/month combination, we have to calculate it or flip forward/backward to the corresponding month. Questions like:

• What day does this year's American Thanksgiving fall on?

• Which months contain a Friday the thirteenth?

• When is July 4th? What day of the week?

• Alternatively, what day of the week is Christmas?

They're hard to figure out until you switch to the right month or look up all the months.

However, mathematically, the answers to these questions or any question that requires matching the day of the week with the day/month combination in a year are predictable, basic, and easy to work out. If you use this one-page calendar instead of a 12-month calendar, it lasts the whole year and is easy to alter for future years. Let me explain.

The 2023 one-page calendar is above. The days of the month are on the lower left, which works for all months if you know that:

• There are 31 days in January, March, May, July, August, October, and December.

• All of the months of April, June, September, and November have 30 days.

• And depending on the year, February has either 28 days (in non-leap years) or 29 days (in leap years).

If you know this, this calendar makes it easy to match the day/month of the year to the weekday.

Here are some instances. American Thanksgiving is always on the fourth Thursday of November. You'll always know the month and day of the week, but the date—the day in November—changes each year.

On any other calendar, you'd have to flip to November to see when the fourth Thursday is. This one-page calendar only requires:

• pick the month of November in the top-right corner to begin.

• drag your finger down until Thursday appears,

• then turn left and follow the monthly calendar until you reach the fourth Thursday.

It's obvious: 2023 is the 23rd American Thanksgiving. For every month and day-of-the-week combination, start at the month, drag your finger down to the desired day, and then move to the left to see which dates match.

What if you knew the day of the week and the date of the month, but not the month(s)?

A different method using the same one-page calendar gives the answer. Which months have Friday the 13th this year? Just:

• begin on the 13th of the month, the day you know you desire,

• then swipe right with your finger till Friday appears.

• and then work your way up until you can determine which months the specific Friday the 13th falls under.

One Friday the 13th occurred in January 2023, and another will occur in October.

The most typical reason to consult a calendar is when you know the month/day combination but not the day of the week.

Compared to single-month calendars, the one-page calendar excels here. Take July 4th, for instance. Find the weekday here:

• beginning on the left on the fourth of the month, as you are aware,

• also begin with July, the month of the year you are most familiar with, at the upper right,

• you should move your two fingers in the opposite directions till they meet: on a Tuesday in 2023.

That's how you find your selected day/month combination's weekday.

Another example: Christmas. Christmas Day is always December 25th, however unless your conventional calendar is open to December of your particular year, a question like "what day of the week is Christmas?" difficult to answer.

Unlike the one-page calendar!

Remember the left-hand day of the month. Top-right, you see the month. Put two fingers, one from each hand, on the date (25th) and the month (December). Slide the day hand to the right and the month hand downwards until they touch.

They meet on Monday—December 25, 2023.

For 2023, that's fine, but what happens in 2024? Even worse, what if we want to know the day-of-the-week/day/month combo many years from now?

I think the one-page calendar shines here.

Except for the blue months in the upper-right corner of the one-page calendar, everything is the same year after year. The months also change in a consistent fashion.

Each non-leap year has 365 days—one more than a full 52 weeks (which is 364). Since January 1, 2023 began on a Sunday and 2023 has 365 days, we immediately know that December 31, 2023 will conclude on a Sunday (which you can confirm using the one-page calendar) and that January 1, 2024 will begin on a Monday. Then, reorder the months for 2024, taking in mind that February will have 29 days in a leap year.

Please note the differences between 2023 and 2024 month placement. In 2023:

• October and January began on the same day of the week.

• On the following Monday of the week, May began.

• August started on the next day,

• then the next weekday marked the start of February, March, and November, respectively.

• Unlike June, which starts the following weekday,

• While September and December start on the following day of the week,

• Lastly, April and July start one extra day later.

Since 2024 is a leap year, February has 29 days, disrupting the rhythm. Month placements change to:

• The first day of the week in January, April, and July is the same.

• October will begin the following day.

• Possibly starting the next weekday,

• February and August start on the next weekday,

• beginning on the following day of the week between March and November,

• beginning the following weekday in June,

• and commencing one more day of the week after that, September and December.

Due to the 366-day leap year, 2025 will start two days later than 2024 on January 1st.

Now, looking at the 2025 calendar, you can see that the 2023 pattern of which months start on which days is repeated! The sole variation is a shift of three days-of-the-week ahead because 2023 had one more day (365) than 52 full weeks (364), and 2024 had two more days (366). Again,

• On Wednesday this time, January and October begin on the same day of the week.

• Although May begins on Thursday,

• August begins this Friday.

• March, November, and February all begin on a Saturday.

• Beginning on a Sunday in June

• Beginning on Monday are September and December,

• and on Tuesday, April and July begin.

In 2026 and 2027, the year will commence on a Thursday and a Friday, respectively.

We must return to our leap year monthly arrangement in 2028. Yes, January 1, 2028 begins on a Saturday, but February, which begins on a Tuesday three days before January, will have 29 days. Thus:

• Start dates for January, April, and July are all Saturdays.

• Given that October began on Sunday,

• Although May starts on a Monday,

• beginning on a Tuesday in February and August,

• Beginning on a Wednesday in March and November,

• Beginning on Thursday, June

• and Friday marks the start of September and December.

This is great because there are only 14 calendar configurations: one for each of the seven non-leap years where January 1st begins on each of the seven days of the week, and one for each of the seven leap years where it begins on each day of the week.

The 2023 calendar will function in 2034, 2045, 2051, 2062, 2073, 2079, 2090, 2102, 2113, and 2119. Except when passing over a non-leap year that ends in 00, like 2100, the repeat time always extends to 12 years or shortens to an extra 6 years.

• The pattern is repeated in 2025's calendar in 2031, 2042, 2053, 2059, 2070, 2081, 2087, 2098, 2110, and 2121.

• The extra 6-year repeat at the end of the century on the calendar for 2026 will occur in the years 2037, 2043, 2054, 2065, 2071, 2082, 2093, 2099, 2105, and 2122.

• The 2027s calendar repeats in 2038, 2049, 2055, 2066, 2077, 2083, 2094, 2100, 2106, and 2117, almost exactly matching the 2026s pattern.

For leap years, the recurrence pattern is every 28 years when not passing a non-leap year ending in 00, or 12 or 40 years when we do. 2024's calendar repeats in 2052, 2080, 2120, 2148, 2176, and 2216; 2028's in 2056, 2084, 2124, 2152, 2180, and 2220.

Knowing January 1st and whether it's a leap year lets you construct a one-page calendar for any year. Try it—you might find it easier than any other alternative!

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.