More on Entrepreneurship/Creators
Jayden Levitt
3 years ago
Billionaire who was disgraced lost his wealth more quickly than anyone in history
If you're not genuine, you'll be revealed.
Sam Bankman-Fried (SBF) was called the Cryptocurrency Warren Buffet.
No wonder.
SBF's trading expertise, Blockchain knowledge, and ability to construct FTX attracted mainstream investors.
He had a fantastic worldview, donating much of his riches to charity.
As the onion layers peel back, it's clear he wasn't the altruistic media figure he portrayed.
SBF's mistakes were disastrous.
Customer deposits were traded and borrowed by him.
With ten other employees, he shared a $40 million mansion where they all had polyamorous relationships.
Tone-deaf and wasteful marketing expenditures, such as the $200 million spent to change the name of the Miami Heat stadium to the FTX Arena
Democrats received a $40 million campaign gift.
And now there seems to be no regret.
FTX was a 32-billion-dollar cryptocurrency exchange.
It went bankrupt practically overnight.
SBF, FTX's creator, exploited client funds to leverage trade.
FTX had $1 billion in customer withdrawal reserves against $9 billion in liabilities in sister business Alameda Research.
Bloomberg Billionaire Index says it's the largest and fastest net worth loss in history.
It gets worse.
SBF's net worth is $900 Million, however he must still finalize FTX's bankruptcy.
SBF's arrest in the Bahamas and SEC inquiry followed news that his cryptocurrency exchange had crashed, losing billions in customer deposits.
A journalist contacted him on Twitter D.M., and their exchange is telling.
His ideas are revealed.
Kelsey Piper says they didn't expect him to answer because people under investigation don't comment.
Bankman-Fried wanted to communicate, and the interaction shows he has little remorse.
SBF talks honestly about FTX gaming customers' money and insults his competition.
Reporter Kelsey Piper was outraged by what he said and felt the mistakes SBF says plague him didn't evident in the messages.
Before FTX's crash, SBF was a poster child for Cryptocurrency regulation and avoided criticizing U.S. regulators.
He tells Piper that his lobbying is just excellent PR.
It shows his genuine views and supports cynics' opinions that his attempts to win over U.S. authorities were good for his image rather than Crypto.
SBF’s responses are in Grey, and Pipers are in Blue.
It's unclear if SBF cut corners for his gain. In their Twitter exchange, Piper revisits an interview question about ethics.
SBF says, "All the foolish sh*t I said"
SBF claims FTX has never invested customer monies.
Piper challenged him on Twitter.
While he insisted FTX didn't use customer deposits, he said sibling business Alameda borrowed too much from FTX's balance sheet.
He did, basically.
When consumers tried to withdraw money, FTX was short.
SBF thought Alameda had enough money to cover FTX customers' withdrawals, but life sneaks up on you.
SBF believes most exchanges have done something similar to FTX, but they haven't had a bank run (a bunch of people all wanting to get their deposits out at the same time).
SBF believes he shouldn't have consented to the bankruptcy and kept attempting to raise more money because withdrawals would be open in a month with clients whole.
If additional money came in, he needed $8 billion to bridge the creditors' deficit, and there aren't many corporations with $8 billion to spare.
Once clients feel protected, they will continue to leave their assets on the exchange, according to one idea.
Kevin OLeary, a world-renowned hedge fund manager, says not all investors will walk through the open gate once the company is safe, therefore the $8 Billion wasn't needed immediately.
SBF claims the bankruptcy was his biggest error because he could have accumulated more capital.
Final Reflections
Sam Bankman-Fried, 30, became the world's youngest billionaire in four years.
Never listen to what people say about investing; watch what they do.
SBF is a trader who gets wrecked occasionally.
Ten first-time entrepreneurs ran FTX, screwing each other with no risk management.
It prevents opposing or challenging perspectives and echo chamber highs.
Twitter D.M. conversation with a journalist is the final nail.
He lacks an experienced crew.
This event will surely speed up much-needed regulation.
It's also prompted cryptocurrency exchanges to offer proof of reserves to calm customers.
SAHIL SAPRU
3 years ago
How I grew my business to a $5 million annual recurring revenue
Scaling your startup requires answering customer demands, not growth tricks.
I cofounded Freedo Rentals in 2019. I reached 50 lakh+ ARR in 6 months before quitting owing to the epidemic.
Freedo aimed to solve 2 customer pain points:
Users lacked a reliable last-mile transportation option.
The amount that Auto walas charge for unmetered services
Solution?
Effectively simple.
Build ports at high-demand spots (colleges, residential societies, metros). Electric ride-sharing can meet demand.
We had many problems scaling. I'll explain using the AARRR model.
Brand unfamiliarity or a novel product offering were the problems with awareness. Nobody knew what Freedo was or what it did.
Problem with awareness: Content and advertisements did a poor job of communicating the task at hand. The advertisements clashed with the white-collar part because they were too cheesy.
Retention Issue: We encountered issues, indicating that the product was insufficient. Problems with keyless entry, creating bills, stealing helmets, etc.
Retention/Revenue Issue: Costly compared to established rivals. Shared cars were 1/3 of our cost.
Referral Issue: Missing the opportunity to seize the AHA moment. After the ride, nobody remembered us.
Once you know where you're struggling with AARRR, iterative solutions are usually best.
Once you have nailed the AARRR model, most startups use paid channels to scale. This dependence, on paid channels, increases with scale unless you crack your organic/inbound game.
Over-index growth loops. Growth loops increase inflow and customers as you scale.
When considering growth, ask yourself:
Who is the solution's ICP (Ideal Customer Profile)? (To whom are you selling)
What are the most important messages I should convey to customers? (This is an A/B test.)
Which marketing channels ought I prioritize? (Conduct analysis based on the startup's maturity/stage.)
Choose the important metrics to monitor for your AARRR funnel (not all metrics are equal)
Identify the Flywheel effect's growth loops (inertia matters)
My biggest mistakes:
not paying attention to consumer comments or satisfaction. It is the main cause of problems with referrals, retention, and acquisition for startups. Beyond your NPS, you should consider second-order consequences.
The tasks at hand should be quite clear.
Here's my scaling equation:
Growth = A x B x C
A = Funnel top (Traffic)
B = Product Valuation (Solving a real pain point)
C = Aha! (Emotional response)
Freedo's A, B, and C created a unique offering.
Freedo’s ABC:
A — Working or Studying population in NCR
B — Electric Vehicles provide last-mile mobility as a clean and affordable solution
C — One click booking with a no-noise scooter
Final outcome:
FWe scaled Freedo to Rs. 50 lakh MRR and were growing 60% month on month till the pandemic ceased our growth story.
How we did it?
We tried ambassadors and coupons. WhatsApp was our most successful A/B test.
We grew widespread adoption through college and society WhatsApp groups. We requested users for referrals in community groups.
What worked for us won't work for others. This scale underwent many revisions.
Every firm is different, thus you must know your customers. Needs to determine which channel to prioritize and when.
Users desired a safe, time-bound means to get there.
This (not mine) growth framework helped me a lot. You should follow suit.
Sammy Abdullah
3 years ago
SaaS payback period data
It's ok and even desired to be unprofitable if you're gaining revenue at a reasonable cost and have 100%+ net dollar retention, meaning you never lose customers and expand them. To estimate the acceptable cost of new SaaS revenue, we compare new revenue to operating loss and payback period. If you pay back the customer acquisition cost in 1.5 years and never lose them (100%+ NDR), you're doing well.
To evaluate payback period, we compared new revenue to net operating loss for the last 73 SaaS companies to IPO since October 2017. (55 out of 73). Here's the data. 1/(new revenue/operating loss) equals payback period. New revenue/operating loss equals cost of new revenue.
Payback averages a year. 55 SaaS companies that weren't profitable at IPO got a 1-year payback. Outstanding. If you pay for a customer in a year and never lose them (100%+ NDR), you're establishing a valuable business. The average was 1.3 years, which is within the 1.5-year range.
New revenue costs $0.96 on average. These SaaS companies lost $0.96 every $1 of new revenue last year. Again, impressive. Average new revenue per operating loss was $1.59.
Loss-in-operations definition. Operating loss revenue COGS S&M R&D G&A (technical point: be sure to use the absolute value of operating loss). It's wrong to only consider S&M costs and ignore other business costs. Operating loss and new revenue are measured over one year to eliminate seasonality.
Operating losses are desirable if you never lose a customer and have a quick payback period, especially when SaaS enterprises are valued on ARR. The payback period should be under 1.5 years, the cost of new income < $1, and net dollar retention 100%.
You might also like
Will Lockett
3 years ago
Tesla recently disclosed its greatest secret.
The VP has revealed a secret that should frighten the rest of the EV world.
Tesla led the EV revolution. Elon Musk's invention offers a viable alternative to gas-guzzlers. Tesla has lost ground in recent years. VW, BMW, Mercedes, and Ford offer EVs with similar ranges, charging speeds, performance, and cost. Tesla's next-generation 4680 battery pack, Roadster, Cybertruck, and Semi were all delayed. CATL offers superior batteries than the 4680. Martin Viecha, Tesla's Vice President, recently told Business Insider something that startled the EV world and will establish Tesla as the EV king.
Viecha mentioned that Tesla's production costs have dropped 57% since 2017. This isn't due to cheaper batteries or devices like Model 3. No, this is due to amazing factory efficiency gains.
Musk wasn't crazy to want a nearly 100% automated production line, and Tesla's strategy of sticking with one model and improving it has paid off. Others change models every several years. This implies they must spend on new R&D, set up factories, and modernize service and parts systems. All of this costs a ton of money and prevents them from refining production to cut expenses.
Meanwhile, Tesla updates its vehicles progressively. Everything from the backseats to the screen has been enhanced in a 2022 Model 3. Tesla can refine, standardize, and cheaply produce every part without changing the production line.
In 2017, Tesla's automobile production averaged $84,000. In 2022, it'll be $36,000.
Mr. Viecha also claimed that new factories in Shanghai and Berlin will be significantly cheaper to operate once fully operating.
Tesla's hand is visible. Tesla selling $36,000 cars for $60,000 This barely beats the competition. Model Y long-range costs just over $60,000. Tesla makes $24,000+ every sale, giving it a 40% profit margin, one of the best in the auto business.
VW I.D4 costs about the same but makes no profit. Tesla's rivals face similar challenges. Their EVs make little or no profit.
Tesla costs the same as other EVs, but they're in a different league.
But don't forget that the battery pack accounts for 40% of an EV's cost. Tesla may soon fully utilize its 4680 battery pack.
The 4680 battery pack has larger cells and a unique internal design. This means fewer cells are needed for a car, making it cheaper to assemble and produce (per kWh). Energy density and charge speeds increase slightly.
Tesla underestimated the difficulty of making this revolutionary new cell. Each time they try to scale up production, quality drops and rejected cells rise.
Tesla recently installed this battery pack in Model Ys and is scaling production. If they succeed, Tesla battery prices will plummet.
Tesla's Model Ys 2170 battery costs $11,000. The same size pack with 4680 cells costs $3,400 less. Once scaled, it could be $5,500 (50%) less. The 4680 battery pack could reduce Tesla production costs by 20%.
With these cost savings, Tesla could sell Model Ys for $40,000 while still making a profit. They could offer a $25,000 car.
Even with new battery technology, it seems like other manufacturers will struggle to make EVs profitable.
Teslas cost about the same as competitors, so don't be fooled. Behind the scenes, they're still years ahead, and the 4680 battery pack and new factories will only increase that lead. Musk faces a first. He could sell Teslas at current prices and make billions while other manufacturers struggle. Or, he could massively undercut everyone and crush the competition once and for all. Tesla and Elon win.
Vitalik
3 years ago
Fairness alternatives to selling below market clearing prices (or community sentiment, or fun)
When a seller has a limited supply of an item in high (or uncertain and possibly high) demand, they frequently set a price far below what "the market will bear." As a result, the item sells out quickly, with lucky buyers being those who tried to buy first. This has happened in the Ethereum ecosystem, particularly with NFT sales and token sales/ICOs. But this phenomenon is much older; concerts and restaurants frequently make similar choices, resulting in fast sell-outs or long lines.
Why do sellers do this? Economists have long wondered. A seller should sell at the market-clearing price if the amount buyers are willing to buy exactly equals the amount the seller has to sell. If the seller is unsure of the market-clearing price, they should sell at auction and let the market decide. So, if you want to sell something below market value, don't do it. It will hurt your sales and it will hurt your customers. The competitions created by non-price-based allocation mechanisms can sometimes have negative externalities that harm third parties, as we will see.
However, the prevalence of below-market-clearing pricing suggests that sellers do it for good reason. And indeed, as decades of research into this topic has shown, there often are. So, is it possible to achieve the same goals with less unfairness, inefficiency, and harm?
Selling at below market-clearing prices has large inefficiencies and negative externalities
An item that is sold at market value or at an auction allows someone who really wants it to pay the high price or bid high in the auction. So, if a seller sells an item below market value, some people will get it and others won't. But the mechanism deciding who gets the item isn't random, and it's not always well correlated with participant desire. It's not always about being the fastest at clicking buttons. Sometimes it means waking up at 2 a.m. (but 11 p.m. or even 2 p.m. elsewhere). Sometimes it's just a "auction by other means" that's more chaotic, less efficient, and has far more negative externalities.
There are many examples of this in the Ethereum ecosystem. Let's start with the 2017 ICO craze. For example, an ICO project would set the price of the token and a hard maximum for how many tokens they are willing to sell, and the sale would start automatically at some point in time. The sale ends when the cap is reached.
So what? In practice, these sales often ended in 30 seconds or less. Everyone would start sending transactions in as soon as (or just before) the sale started, offering higher and higher fees to encourage miners to include their transaction first. Instead of the token seller receiving revenue, miners receive it, and the sale prices out all other applications on-chain.
The most expensive transaction in the BAT sale set a fee of 580,000 gwei, paying a fee of $6,600 to get included in the sale.
Many ICOs after that tried various strategies to avoid these gas price auctions; one ICO notably had a smart contract that checked the transaction's gasprice and rejected it if it exceeded 50 gwei. But that didn't solve the issue. Buyers hoping to game the system sent many transactions hoping one would get through. An auction by another name, clogging the chain even more.
ICOs have recently lost popularity, but NFTs and NFT sales have risen in popularity. But the NFT space didn't learn from 2017; they do fixed-quantity sales just like ICOs (eg. see the mint function on lines 97-108 of this contract here). So what?
That's not the worst; some NFT sales have caused gas price spikes of up to 2000 gwei.
High gas prices from users fighting to get in first by sending higher and higher transaction fees. An auction renamed, pricing out all other applications on-chain for 15 minutes.
So why do sellers sometimes sell below market price?
Selling below market value is nothing new, and many articles, papers, and podcasts have written (and sometimes bitterly complained) about the unwillingness to use auctions or set prices to market-clearing levels.
Many of the arguments are the same for both blockchain (NFTs and ICOs) and non-blockchain examples (popular restaurants and concerts). Fairness and the desire not to exclude the poor, lose fans or create tension by being perceived as greedy are major concerns. The 1986 paper by Kahneman, Knetsch, and Thaler explains how fairness and greed can influence these decisions. I recall that the desire to avoid perceptions of greed was also a major factor in discouraging the use of auction-like mechanisms in 2017.
Aside from fairness concerns, there is the argument that selling out and long lines create a sense of popularity and prestige, making the product more appealing to others. Long lines should have the same effect as high prices in a rational actor model, but this is not the case in reality. This applies to ICOs and NFTs as well as restaurants. Aside from increasing marketing value, some people find the game of grabbing a limited set of opportunities first before everyone else is quite entertaining.
But there are some blockchain-specific factors. One argument for selling ICO tokens below market value (and one that persuaded the OmiseGo team to adopt their capped sale strategy) is community dynamics. The first rule of community sentiment management is to encourage price increases. People are happy if they are "in the green." If the price drops below what the community members paid, they are unhappy and start calling you a scammer, possibly causing a social media cascade where everyone calls you a scammer.
This effect can only be avoided by pricing low enough that post-launch market prices will almost certainly be higher. But how do you do this without creating a rush for the gates that leads to an auction?
Interesting solutions
It's 2021. We have a blockchain. The blockchain is home to a powerful decentralized finance ecosystem, as well as a rapidly expanding set of non-financial tools. The blockchain also allows us to reset social norms. Where decades of economists yelling about "efficiency" failed, blockchains may be able to legitimize new uses of mechanism design. If we could use our more advanced tools to create an approach that more directly solves the problems, with fewer side effects, wouldn't that be better than fiddling with a coarse-grained one-dimensional strategy space of selling at market price versus below market price?
Begin with the goals. We'll try to cover ICOs, NFTs, and conference tickets (really a type of NFT) all at the same time.
1. Fairness: don't completely exclude low-income people from participation; give them a chance. The goal of token sales is to avoid high initial wealth concentration and have a larger and more diverse initial token holder community.
2. Don’t create races: Avoid situations where many people rush to do the same thing and only a few get in (this is the type of situation that leads to the horrible auctions-by-another-name that we saw above).
3. Don't require precise market knowledge: the mechanism should work even if the seller has no idea how much demand exists.
4. Fun: The process of participating in the sale should be fun and game-like, but not frustrating.
5. Give buyers positive expected returns: in the case of a token (or an NFT), buyers should expect price increases rather than decreases. This requires selling below market value.
Let's start with (1). From Ethereum's perspective, there is a simple solution. Use a tool designed for the job: proof of personhood protocols! Here's one quick idea:
Mechanism 1 Each participant (verified by ID) can buy up to ‘’X’’ tokens at price P, with the option to buy more at an auction.
With the per-person mechanism, buyers can get positive expected returns for the portion sold through the per-person mechanism, and the auction part does not require sellers to understand demand levels. Is it race-free? The number of participants buying through the per-person pool appears to be high. But what if the per-person pool isn't big enough to accommodate everyone?
Make the per-person allocation amount dynamic.
Mechanism 2 Each participant can deposit up to X tokens into a smart contract to declare interest. Last but not least, each buyer receives min(X, N / buyers) tokens, where N is the total sold through the per-person pool (some other amount can also be sold by auction). The buyer gets their deposit back if it exceeds the amount needed to buy their allocation.
No longer is there a race condition based on the number of buyers per person. No matter how high the demand, it's always better to join sooner rather than later.
Here's another idea if you like clever game mechanics with fancy quadratic formulas.
Mechanism 3 Each participant can buy X units at a price P X 2 up to a maximum of C tokens per buyer. C starts low and gradually increases until enough units are sold.
The quantity allocated to each buyer is theoretically optimal, though post-sale transfers will degrade this optimality over time. Mechanisms 2 and 3 appear to meet all of the above objectives. They're not perfect, but they're good starting points.
One more issue. For fixed and limited supply NFTs, the equilibrium purchased quantity per participant may be fractional (in mechanism 2, number of buyers > N, and in mechanism 3, setting C = 1 may already lead to over-subscription). With fractional sales, you can offer lottery tickets: if there are N items available, you have a chance of N/number of buyers of getting the item, otherwise you get a refund. For a conference, groups could bundle their lottery tickets to guarantee a win or a loss. The certainty of getting the item can be auctioned.
The bottom tier of "sponsorships" can be used to sell conference tickets at market rate. You may end up with a sponsor board full of people's faces, but is that okay? After all, John Lilic was on EthCC's sponsor board!
Simply put, if you want to be reliably fair to people, you need an input that explicitly measures people. Authentication protocols do this (and if desired can be combined with zero knowledge proofs to ensure privacy). So we should combine the efficiency of market and auction-based pricing with the equality of proof of personhood mechanics.
Answers to possible questions
Q: Won't people who don't care about your project buy the item and immediately resell it?
A: Not at first. Meta-games take time to appear in practice. If they do, making them untradeable for a while may help mitigate the damage. Using your face to claim that your previous account was hacked and that your identity, including everything in it, should be moved to another account works because proof-of-personhood identities are untradeable.
Q: What if I want to make my item available to a specific community?
A: Instead of ID, use proof of participation tokens linked to community events. Another option, also serving egalitarian and gamification purposes, is to encrypt items within publicly available puzzle solutions.
Q: How do we know they'll accept? Strange new mechanisms have previously been resisted.
A: Having economists write screeds about how they "should" accept a new mechanism that they find strange is difficult (or even "equity"). However, abrupt changes in context effectively reset people's expectations. So the blockchain space is the best place to try this. You could wait for the "metaverse", but it's possible that the best version will run on Ethereum anyway, so start now.
Dmitrii Eliuseev
2 years ago
Creating Images on Your Local PC Using Stable Diffusion AI
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 condaInstall 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 --upgradeDownload 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 1Almost. 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 1Stable 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 1The 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 0.8It 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 ldmHugging 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.ckptThis 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.