Black Forest Labs – Flux1.0 / 1.1 image generation

Background

Black Forest Labs, a German based company, released FLUX on the 1st August this year. It was released without the usual build-up and hype that has sometimes accompanied other Ai launches. Flux is however a real game changer.

Without doubt its launch put it head and shoulders above any of the competition for creating images, and already an improved version has just been released. Flux1.1 addressed limitations in user control, speed, and output refinement, meaning that the output more closely matches the text prompts, and is quicker. In Ai land, quicker usually also means less expensive.

Why isn’t Flux a standalone application?

Flux can’t be downloaded directly onto a standard computer and run with an easy to use interface. Like many other Ai developments, it requires a platform to run on. The reason for this is that Flux is the engine to make things work, and isn’t intended as a standalone system. For example Flux is integrated into xAi’s Grok-2, Elton Musk’s rival to Chat-GPT, giving Grok an image generator in much the same way that Dall-E is part of OpenAi’s Chat GPT.

The platform’s task is to act as an intermediary handling all the technical stuff; creating links to the servers; setting up the APIs, and so on, such that users only have to manage a simplified front end without the need to dig deep into the system or learn anything technical.

A simple analogy is this; imagine Flux to be like a car engine – it’s very high performance – but it’s still only the engine. It requires a vehicle body and some custom tuning to use it effectively.

And that car body takes the form of a variety of different platforms – of which there are many different options. Best known are HuggingFace, Replicate, AWS (Amazon SageMaker), and Google Cloud AI (Vertex AI), but there are many more.

To stick with the car analogy, Hugging Face is probably like a comfortable SUV, Replicate like a sports car, and both Amazon and Google very heavy duty trucks with giant wheels.

A heavy duty truck is not the kind of vehicle that’s of much use to us small video makers – there’s a steep learning curve to control them, they are expensive to run for small jobs, and parking at the supermarket can be a problem!

Part of the added value of these platforms is to tune the engine to some degree, so Flux outputs can be slightly different in the same way that engines might be tuned to perform differently in different vehicles.

Replicate’s interface is really easy to use, and it was my choice of platform for Flux.

 It’s straightforward, has good guides, and produces the type of output that I want. Whilst there is some flexibility in ‘tuning’ the style and type images, I’m very happy with those that it produces without needing any extra tweaks. And I know that it still has the potential to allow me to create my own fine-tuning if I need to later.

How to easily produce Flux images using Replicate.com

This is a very easy (non technical) guide to get started with any of the Flux models, and at the same time get access to many other utilities and Ai tools that are available on Replicate.

Firstly open replicate.com,

 click ‘Get started’, and then sign in to GitHub.

If you don’t have an account, simply create one:

You’ll need your email, create a password, create a username, choose email preference option, and finally verify your account. Then you are all set and will be returned to the Replicate page.

Whilst it’s possible to use Replicate without payment (some of the options are free), if you want to do anything useful you’ll need to add a credit card to the account. You’ll be prompted to add a payment method.

How much does it cost and how am I invoiced?

Each month you will only pay for what you use, so it’s not a standing subscription. If you don’t generate any output there is no charge. You are also encouraged to put a cap on your account to avoid any possible overspending. Mine is set at a miserly $5/month, and so far it’s been more than sufficient for the several hundred images I’ve produced!

In replicate, you’ll see many other programs and utilities on the platform, don’t be afraid to experiment. The estimated cost per image will appear at the top and bottom of the relevant page. Check it out before using the model you’ve selected.

Which Flux version is best to use to generate images?

Black Forest Labs launched 3 models on 1^st August2024:

• Flux 1.0 Schnell (@ 0.3 cents/image) (basic, but 1/8 of the price of the Dev version)
• Flux 1.0-Dev (@ 2.5 cents/image)
• Flux1.0-Pro (@ 5.5 cents/image)
• Flux 1.1 Pro was added in October 2024 @ 4 cents/image.

My preference is for Flux 1.0 Dev. It suits my purposes and isn’t too pricy, but test out alternatives to find your own preferred output quality. Simply select the option from the range available on the ‘Explore’ Tab.

How to generate an image using Flux1.0 dev on Replicate

It’s really easy to generate an image.

Under ‘explore’ tab, look for ‘Flux1-Dev’ and click on it.

This takes you to this input page.

To start generating there are only 4 actions required:

1. Enter your prompt – it can be something very simple – ‘spotted fluffy cat, riding a skateboard watched by lots of spectators. Colorful modern 3d animation style’.
2. Select the image size – 16:9 is widescreen HD
3. Select the image output – It defaults to webp, so change it to either jpg or png (either are fine)
4. Press ‘RUN’.

When it’s generated, click ‘Download’ to save it locally.

All the other settings can be left at default, ready to play with another time!

(Btw this same prompt was used for another ‘cat on a skateboard’ creation shown below, except the last sentence was replaced with ‘Photo Realistic’)
If you want to go back to see any later, older images will be found under the‘Dashboard’-‘Predictions’ tab.

              Why does it cost?

One thing to be clear about with Ai is that the really good stuff isn’t free! Someone has to pay for the programming, compute, and the power (energy) to drive them. They aren’t simply using the top of the range $1,500 Nvida 4090Ti GPU’s in a powerful PC, instead they might access banks of Nvidia T100 GPU’s at about $10,000 each, or even the much more powerful H100 at about $40,000 each.

How does it work?

Contrary to popular misunderstanding, Ai generated images are not simply copied from the internet and pasted into a composite image. Each image is generated by the Ai model using its understanding of the world and how it works. Because Ai image generation is in its infancy, Ai can sometimes (or even often) get details wrong. For example hands, and fingers in particular, have posed a problem for Ai, and even today they’re not always perfect. As an illustration these fingers produced by Flux1.0 (the image on the right) aren’t quite correct – which might also explain his sorrowful expression. The real challenge will come when they are perfect. Then how will we know what is real?

How can I produce exactly the images that I want?

The key to all text to image is the quality of the prompts. Indeed, it will clearly create a small industry for competent ‘Prompt Writers’.

However, to be honest, I use relatively simple prompts and am happy to accept the variability of the output. Some of my prompts are shown, but even simpler ones such as ‘man using laptop computer on his desk’ produces excellent results.

Dealing with Ai generated output has a lot to do with setting realistic personal expectations of the Ai. Hoping that Ai will produce exactly the image that you hold in your head is a guaranteed way to create massive frustration. Believe me, I’ve been there! The best way is not to hold any preconceived ideas of exactly how it might look, and utilize the output in whatever form it appears.

Ai is still improving, and it’s easy to forget that just two years ago if you said that life-like photo realistic images could be created from simple text prompts you’d be labeled as delusional. Ai will only get better, but for now my advice is to accept what it can do today, rather than fighting it hoping for perfection.

In some cases I’ve simply tweaked my storyline to suit the image output. The creative part for us is to utilize the images produced in ways that can be used to enhance a story in a video (or document such as this) without becoming frustrated that the man was wearing glasses…

Images were generated using Black Forrest Labs Flux1.x