Complex and comprehensive
It has been over two years now since the code for ComfyUI was released on GitHub in early 2023, and over the past year it has become my go-to application for generating images via Stable Diffusion and Flux. While it doesn't always handle every model as well as it could, it has several advantages in terms of speed and the ability to customise to a huge extent. It's precisely that customisation, as well as the fact that not all aspects of checkpoints (such as VAEs) are downloaded automatically with models that cause elements of frustration, but with some dedication to mastering its admittedly complex interface it is the best app for image generation out there.
One thing that has been addressed with the latest version is a desktop installer. Actually released in October, it is still in beta and receives fairly regular updates, including the integration of the ComfyUI Manager which was previously a separate installation - all of that being done via the terminal through cloning the GitHub repository. To be honest, there are some basic housekeeping elements that users should be prepared to engage with (such as updating the default version of Python that seems to be installed even on new Macs, as well as installing Homebrew), but the ability to download and have the program up and running in minutes is a welcome change.
ComfyUI is extremely capable, but remains complex to use for novice users.
How does it work?
Several reviews of the app tend to begin from the standing observation that it is intuitive and easy for beginners to get to grips with. It is not. Not in the slightest. If you have experience of working with IDEs and other programming interfaces, certainly it is intuitive in comparison to that, but I also work with a number of students who while having a degree of technical savvy with programs such as Adobe creative tools are usually stumped completely when they first open up ComfyUI. I personally do find it increasingly intuitive to work with, but that's because I've been using it for over a year and have become extremely familiar with its workflows. If you're new to using Stable Diffusion and similar to create images, then prepare for a steep learning curve. The Ease of Use mark at the end of this review is because it can now be downloaded, along with the Manager, straight to your PC. Without that, this would be 1 out of 5 for simplicity.
That is not actually intended as a criticism of the program. While I would have appreciated a little more handholding when first getting to grips with it, over the past few months, I now fully appreciate the sheer flexibility of its approach. It works via nodes, with users adding base models and checkpoints, loras, samplers, VAEs and the various other aspects that will handle the generation of your image. If such terms as loras and VAEs mean nothing to you it is very much worthwhile starting to get to grips with these and experimenting with them in the interface as they will greatly affect the appearance and style of your image. One of the great things about the program is that, once you have mastered some of the basics, it really does encourage you to experiment and play around with different settings.
Once different elements of the image generation model are loaded, they are connected together via connectors which are colour-coded. These, while appearing overwhelming at first, do actually become much easier to use after only a few encounters in that the red connector for a VAE in a base model or checkpoint is clearly intended to join up the red dot on a VAE decoder node. It does take practice, but if you've ever used applciations such as the database designer in PHPMyAdmin it quickly becomes second nature. Connecting nodes together creates a workflow which passes parameters on to your text encoder where you enter prompts: these are then interpreted and output as an image which you can set initial sizes for directly in the workflow or use a resizer to upscale.
Another important point to make is in terms of the machine you use to create your images. While I have moved almost completely to Macs now and cannot talk about the experience on a PC, there is a notable difference between even the M4 Max Macbook I use and the M2 Max Studio, with some models such as Flux creating images in less than half the time. It is not that they were unusable on the older Studio (in contrast to an M2 Macbook Air, which can only run much smaller Stable Diffusion models), but that generating images in a minute or less tends to encourage the kind of experimentation that would stop when they required five minutes or more. ComfyUI does seem to work more quickly than some of the competition, but this is something where you will definitely notice the benefits of working with a more powerful computer.
Is it effective?
The answer is very. While complete newbies will probably be very confused initially by the interface for ComfyUI, it certainly does provide great advantages to those looking to create AI art. It is not easy, and one thing that could be greatly improved is the community manual which is far from being the most intuitive guide I've come across. Nonetheless, with a willingness to experiment with different nodes - which can easily be saved to their own folder - in only a short time you will come on in leaps and bounds with the app. What's more, because generation information is saved with the image in PNG format, for images you particularly like you can simply drag and drop them in the interface to launch the workflow to be used with future iterations. For a free program, its capabilities are little short of astonishing.