AI images are like funhouse mirrors.

Google's AI Blog introduced the puppy-slug in the summer of 2015.

Puppy-slug isn't a single image or character. "Puppy-slug" refers to Google's DeepDream's unsettling psychedelia. This tool uses convolutional neural networks to train models to recognize dataset entities. If researchers feed the model millions of dog pictures, the network will learn to recognize a dog.

DeepDream used neural networks to analyze and classify image data as well as generate its own images. DeepDream's early examples were created by training a convolutional network on dog images and asking it to add "dog-ness" to other images. The models analyzed images to find dog-like pixels and modified surrounding pixels to highlight them.

Puppy-slugs and other DeepDream images are ugly. Even when they don't trigger my trypophobia, they give me vertigo when my mind tries to reconcile familiar features and forms in unnatural, physically impossible arrangements. I feel like I've been poisoned by a forbidden mushroom or a noxious toad. I'm a Lovecraft character going mad from extradimensional exposure. They're gross!

Is this really how AIs see the world? This is possibly an even more unsettling topic that DeepDream raises than the blatant abjection of the images.

When these photographs originally circulated online, many friends were startled and scandalized. People imagined a computer's imagination would be literal, accurate, and boring. We didn't expect vivid hallucinations and organic-looking formations.

DeepDream's images didn't really show the machines' imaginations, at least not in the way that scared some people. DeepDream displays data visualizations. DeepDream reveals the "black box" of convolutional network training.

Some of these images look scary because the models don't "know" anything, at least not in the way we do.

These images are the result of advanced algorithms and calculators that compare pixel values. They can spot and reproduce trends from training data, but can't interpret it. If so, they'd know dogs have two eyes and one face per head. If machines can think creatively, they're keeping it quiet.

You could be forgiven for thinking otherwise, given OpenAI's Dall-impressive E's results. From a technological perspective, it's incredible.

Arthur C. Clarke once said, "Any sufficiently advanced technology is indistinguishable from magic." Dall-magic E's requires a lot of math, computer science, processing power, and research. OpenAI did a great job, and we should applaud them.

Dall-E and similar tools match words and phrases to image data to train generative models. Matching text to images requires sorting and defining the images. Untold millions of low-wage data entry workers, content creators optimizing images for SEO, and anyone who has used a Captcha to access a website make these decisions. These people could live and die without receiving credit for their work, even though the project wouldn't exist without them.

This technique produces images that are less like paintings and more like mirrors that reflect our own beliefs and ideals back at us, albeit via a very complex prism. Due to the limitations and biases that these models portray, we must exercise caution when viewing these images.

The issue was succinctly articulated by artist Mimi Onuoha in her piece "On Algorithmic Violence":

As we continue to see the rise of algorithms being used for civic, social, and cultural decision-making, it becomes that much more important that we name the reality that we are seeing. Not because it is exceptional, but because it is ubiquitous. Not because it creates new inequities, but because it has the power to cloak and amplify existing ones. Not because it is on the horizon, but because it is already here.

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.

I was entering tens of millions of records per hour when I first published Slonik PostgreSQL client for Node.js. The data being entered was usually flat, making it straightforward to use INSERT INTO ... SELECT * FROM unnset() pattern. I advocated the unnest approach for inserting rows in groups (that was part I).

However, today I’ve found a better way: jsonb_to_recordset.

jsonb_to_recordset expands the top-level JSON array of objects to a set of rows having the composite type defined by an AS clause.

jsonb_to_recordset allows us to query and insert records from arbitrary JSON, like unnest. Since we're giving JSON to PostgreSQL instead of unnest, the final format is more expressive and powerful.

SELECT *
FROM json_to_recordset('[{"name":"John","tags":["foo","bar"]},{"name":"Jane","tags":["baz"]}]')
AS t1(name text, tags text[]);
 name |   tags
------+-----------
 John | {foo,bar}
 Jane | {baz}
(2 rows)

Let’s demonstrate how you would use it to insert data.

Inserting data using json_to_recordset

Say you need to insert a list of people with attributes into the database.

const persons = [
  {
    name: 'John',
    tags: ['foo', 'bar']
  },
  {
    name: 'Jane',
    tags: ['baz']
  }
];

You may be tempted to traverse through the array and insert each record separately, e.g.

for (const person of persons) {
  await pool.query(sql`
    INSERT INTO person (name, tags)
    VALUES (
      ${person.name},
      ${sql.array(person.tags, 'text[]')}
    )
  `);
}

It's easier to read and grasp when working with a few records. If you're like me and troubleshoot a 2M+ insert query per day, batching inserts may be beneficial.

What prompted the search for better alternatives.

Inserting using unnest pattern might look like this:

await pool.query(sql`
  INSERT INTO public.person (name, tags)
  SELECT t1.name, t1.tags::text[]
  FROM unnest(
    ${sql.array(['John', 'Jane'], 'text')},
    ${sql.array(['{foo,bar}', '{baz}'], 'text')}
  ) AS t1.(name, tags);
`);

You must convert arrays into PostgreSQL array strings and provide them as text arguments, which is unsightly. Iterating the array to create slices for each column is likewise unattractive.

However, with jsonb_to_recordset, we can:

await pool.query(sql`
  INSERT INTO person (name, tags)
  SELECT *
  FROM jsonb_to_recordset(${sql.jsonb(persons)}) AS t(name text, tags text[])
`);

In contrast to the unnest approach, using jsonb_to_recordset we can easily insert complex nested data structures, and we can pass the original JSON document to the query without needing to manipulate it.

In terms of performance they are also exactly the same. As such, my current recommendation is to prefer jsonb_to_recordset whenever inserting lots of rows or nested data structures.

5 VC investment criteria

I reviewed 200 startup business concepts last week. Brainache.

The enterprises sold various goods and services. The concepts were achingly similar: give us money, we'll produce a product, then get more to expand. No different from daily plans and pitches.

Most of those 200 plans sounded plausible. But 10% looked venture-worthy. 90% of startups need alternatives to venture finance.

With the success of VC-backed businesses and the growth of venture funds, a common misperception is that investors would fund any decent company idea. Finding investors that believe in the firm and founders is the key to funding.

Incorrect. Venture capital needs investing in certain enterprises. If your startup doesn't match the model, as most early-stage startups don't, you can revise your business plan or locate another source of capital.

Before spending six months pitching angels and VCs, make sure your startup fits these criteria.

Likely to generate $100 million in sales

First, I check the income predictions in a pitch deck. If it doesn't display $100M, don't bother.

The math doesn't work for venture financing in smaller businesses.

Say a fund invests $1 million in a startup valued at $5 million that is later acquired for $20 million. That's a win everyone should celebrate. Most VCs don't care.

Consider a $100M fund. The fund must reach $360M in 7 years with a 20% return. Only 20-30 investments are possible. 90% of the investments will fail, hence the 23 winners must return $100M-$200M apiece. $15M isn't worth the work.

Angel investors and tiny funds use the same ideas as venture funds, but their smaller scale affects the calculations. If a company can support its growth through exit on less than $2M in angel financing, it must have $25M in revenues before large companies will consider acquiring it.

Aiming for Hypergrowth

A startup's size isn't enough. It must expand fast.

Developing a great business takes time. Complex technology must be constructed and tested, a nationwide expansion must be built, or production procedures must go from lab to pilot to factories. These can be enormous, world-changing corporations, but venture investment is difficult.

The normal 10-year venture fund life. Investments are made during first 3–4 years.. 610 years pass between investment and fund dissolution. Funds need their investments to exit within 5 years, 7 at the most, therefore add a safety margin.

Longer exit times reduce ROI. A 2-fold return in a year is excellent. Loss at 2x in 7 years.

Lastly, VCs must prove success to raise their next capital. The 2nd fund is raised from 1st fund portfolio increases. Third fund is raised using 1st fund's cash return. Fund managers must raise new money quickly to keep their jobs.

Branding or technology that is protected

No big firm will buy a startup at a high price if they can produce a competing product for less. Their development teams, consumer base, and sales and marketing channels are large. Who needs you?

Patents, specialist knowledge, or brand name are the only answers. The acquirer buys this, not the thing.

I've heard of several promising startups. It's not a decent investment if there's no exit strategy.

A company that installs EV charging stations in apartments and shopping areas is an example. It's profitable, repeatable, and big. A terrific company. Not a startup.

This building company's operations aren't secret. No technology to protect, no special information competitors can't figure out, no go-to brand name. Despite the immense possibilities, a large construction company would be better off starting their own.

Most venture businesses build products, not services. Services can be profitable but hard to safeguard.

Probable purchase at high multiple

Once a software business proves its value, acquiring it is easy. Pharma and medtech firms have given up on their own research and instead acquire startups after regulatory permission. Many startups, especially in specialized areas, have this weakness.

That doesn't mean any lucrative $25M-plus business won't be acquired. In many businesses, the venture model requires a high exit premium.

A startup invents a new glue. 3M, BASF, Henkel, and others may buy them. Adding more adhesive to their catalogs won't boost commerce. They won't compete to buy the business. They'll only buy a startup at a profitable price. The acquisition price represents a moderate EBITDA multiple.

The company's $100M revenue presumably yields $10m in profits (assuming they’ve reached profitability at all). A $30M-$50M transaction is likely. Not terrible, but not what venture investors want after investing $25M to create a plant and develop the business.

Private equity buys profitable companies for a moderate profit multiple. It's a good exit for entrepreneurs, but not for investors seeking 10x or more what PE firms pay. If a startup offers private equity as an exit, the conversation is over.

Constructed for purchase

The startup wants a high-multiple exit. Unless the company targets $1B in revenue and does an IPO, exit means acquisition.

If they're constructing the business for acquisition or themselves, founders must decide.

If you want an indefinitely-running business, I applaud you. We need more long-term founders. Most successful organizations are founded around consumer demands, not venture capital's urge to grow fast and exit. Not venture funding.

if you don't match the venture model, what to do

VC funds moonshots. The 10% that succeed are extraordinary. Not every firm is a rocketship, and launching the wrong startup into space, even with money, will explode.

But just because your startup won't make $100M in 5 years doesn't mean it's a bad business. Most successful companies don't follow this model. It's not venture capital-friendly.

Although venture capital gets the most attention due to a few spectacular triumphs (and disasters), it's not the only or even most typical option to fund a firm.

Other ways to support your startup:

• Personal and family resources, such as credit cards, second mortgages, and lines of credit

• bootstrapping off of sales

• government funding and honors

• Private equity & project financing

• collaborating with a big business

• Including a business partner

Before pitching angels and VCs, be sure your startup qualifies. If so, include them in your pitch.

“Poor distribution, not poor goods, is the main reason for failure” — Peter Thiel.

Here's an easy way to conceptualize "go-to-market" for your distribution plan.

One equation captures the concept:

Distribution = Ecosystem Participants + Incentives

Draw your customers' ecosystem. Set aside your goods and consider your consumer's environment. Who do they deal with daily? 

1. First, list each participant. You want an exhaustive list, but here are some broad categories.

• In-person media services

• Websites

• Events\Networks

• Financial education and banking

• Shops

• Staff

• Advertisers

• Twitter influencers

1. Draw influence arrows. Who's affected? I'm not just talking about Instagram selfie-posters. Who has access to your consumer and could promote your product if motivated?

The thicker the arrow, the stronger the relationship. Include more "influencers" if needed. Customer ecosystems are complex.

3. Incentivize ecosystem players. “Show me the incentive and I will show you the result.“, says Warren Buffet's business partner Charlie Munger.

Strong distribution strategies encourage others to promote your product to your target market by incentivizing the most prominent players. Incentives can be financial or non-financial.

Financial rewards

Usually, there's money. If you pay Facebook, they'll run your ad. Salespeople close deals for commission. Giving customers bonus credits will encourage referrals.

Most businesses underuse non-financial incentives.

Non-cash incentives

Motivate key influencers without spending money to expand quickly and cheaply. What can you give a client-connector for free?

Here are some ideas:

Are there any other features or services available?

Titles or status? Tinder paid college "ambassadors" for parties to promote its dating service.

Can I get early/free access? Facebook gave a select group of developers "exclusive" early access to their AR platform.

Are you a good host? Pharell performed at YPlan's New York launch party.

Distribution? Apple's iPod earphones are white so others can see them.

Have an interesting story? PR rewards journalists by giving them a compelling story to boost page views.

Prioritize distribution.

More time spent on distribution means more room in your product design and business plan. Once you've identified the key players in your customer's ecosystem, talk to them.

Money isn't your only resource. Creative non-monetary incentives may be more effective and scalable. Give people something useful and easy to deliver.