Since the beginning of the AI hype, Nvidia hardware has left no alternatives when it comes to neural networks, but that has changed. Software developers have finally adapted AI models for AMD Radeon graphics cards. The results are truly amazing, but let’s go through the process.
Attention! The editors of ITC.ua ask that this text be taken solely as a journalistic study of AI capabilities and do not encourage the creation and public dissemination of texts and images that violate generally accepted moral standards We treat all people with respect, regardless of their racial, religious, or sexual identity.
Content
Tools
In the experiment, we used the software Amuse 2.1which is perfectly adapted to work with AMD hardware, has a user-friendly interface and is suitable for both beginners and professionals. This software allows you to generate images from text descriptions and more.
We also installed a text chatbot. Program LM Studio operates with text neural networks like ChatGPT. The utility is extremely functional, as it can work with a processor or video card and has a large library of language models (including uncensored ones).
We tested neural networks on a computer with this configuration:
- AMD Ryzen 9 7900 PBO 150 W processor with MSI MEG S360 cooling;
- Video card AMD Radeon RX 7900 XTX 24 GB in the reference design;
- Memory 32 GB DDR5 6000CL30 + 1 TB SSD PCI-E 4.0 from Kingston;
Installing and configuring Amuse
This point could be skipped because the installation process is incredibly simple and consists of several steps:
- Download the installer Amuse from the official website;
- Install it on a computer with Windows OS;
- Launch the program and start working;
After launching, the utility will ask you to download the trained models (they take up a lot of space). In total, three models for each of the intensity modes are required to generate images according to the description.
Hardware acceleration works by default, i.e. the neural network loads the video card, not the CPU. In the «Fast» mode, images are fired at lightning speed (up to 10 images in 10 seconds), but their quality often leaves much to be desired. The load on the video card is low.
Create masterpieces without limits
The first tab allows you to generate images based on a text description. You can write it in Ukrainian, but it is recommended to write it in English for greater detail. The description should be as simple, concise, and clear as possible for the machine.
You can queue up to 10 images in three formats at once: square, horizontal, and vertical. The resolution is 1216×832 pixels, no more.
The second tab allows you to apply filters to the target image. In a few seconds, you can create a Disney, Cyberpunk, or Neo from the Matrix-style photo of yourself. It all depends on your imagination, as described in the text description, and the selected preset.
The «Design» tab is the most interesting because it can turn a thumbnail into a full-fledged image. It’s enough to draw a little man made of several sticks with the mouse, set a description, and as a result, you get a warrior that looks almost exactly as you imagined.
Not satisfied? Generate a new image in three seconds. The search method always manages to get something that fits. Yes, there are apps for iPad with similar functionality, but Amuse does it all for free without any subscriptions.
Censorship is present in Amuse 2.1 if you use standard models or those available in the official catalog. Those who want to generate hentai have to search for trained models on the Internet on their own.
However, there are a few nuances worth knowing about. For example, the neuron sometimes agrees to draw trash, such as people who eat cats, but censors other images. We barely managed to get Amuse to draw a blonde in a bikini, no more.
So, with standard models, you can’t draw fully or partially naked women, the images will be blurry, but this rule breaks down when you ask to draw a transgender man. For some reason, the software draws them, and immediately from genitals. We won’t show these explicit photos here, but you can see them on the ITC.ua Telegram channel.
It’s a bit strange that this trained model thinks that a woman without a bikini should not be drawn. But a muscular man who identifies as a woman, has long wavy hair, wears shoes with 18-centimeter heels and no clothes, he has no problem depicting.
These are strange standards, and there is nothing you can do about it. But with other models, you can bypass any restrictions and create the content you want, without censorship or tolerance.
Get a similar result with online image generators never will not work because they all have many limitations and are «stifled by» developers. This, in my opinion, is the main advantage of local neural networks – absolute (or almost absolute) permissiveness.
Installing and configuring LM Studio
As in the first case, the installation process is as simple as possible. The user just needs to download LM Studio with AMD ROCm from the official website. After installation, the program will offer to download the required models.
Note that you need a lot of space, a lot of it, so it’s better to choose the desired location on the drive in advance so as not to fill the system disk 100%. During testing, we downloaded all the models for Amuse and LM Studio to 200 GB!
The program interface is not difficult to understand -the most important thing is to download the desired model from the recommended list or using a manual search. Set the maximum performance mode of the video card in the column on the right side and start working.
The difference between the video card and the processor is huge, so be sure to prioritize the GPU if you want the neural network to work faster. Please note that some models require 16 GB of video memory, although on average, 8 GB is enough. The AMD Radeon RX 7600 XT 16 GB is a perfect fit for such tasks with a huge margin.
Otherwise, get ready for significant RAM consumption. It is recommended to have at least 32 GB of RAM, and preferably 64 GB of RAM for comfortable work with several models simultaneously. In this case, the response speed will decrease compared to GPU operation.
Chat without censorship
Standard textual models do not allow making offensive queries in cases where it comes to racial, religious, sexual identification of people, etc. Any cloud-based neurons, such as ChatGPT, Copilot, etc., behave the same way.
With the RX 7900 XTX, the base models are fast, deliver average quality results (ChatGPT 4.0 is better), and it all works locally without any subscriptions or the Internet.
Not all models are the same, so you will have to experiment to achieve the desired result. Some models take up 10+ GB of space and require up to 16 GB of memory (RAM or video).
But the main question remains unanswered: is it possible to create your own neuron with «vodka and» night butterflies? Yes, it is, but you will have to download uncensored models.
In fact, these are properly trained models that do not constrain your computer, allowing artificial intelligence to answer any questions. Where you can get a clear, understandable and precise answer from LM Studio, often with confirmed facts, proof, etc., and ChatGPT will immediately end the conversation.
Similarly, you can ask to tell an intolerant joke, or write an essay about how insecure student Mikhail did God knows what with teacher Zhanna after a couple in the toilet, or anything else that comes to mind.
In general, the quality of the generated content is not bad (sometimes bad), but it will have to be finalized, i.e., some minor gaffes, inconsistencies, etc. The worst texts are those on technical topics with a lot of precise data. For example, it is difficult to describe technology with such a neuron because there are almost half of the mistakes when it comes to technical specifications.
The basic models are mediocre at keeping context. ChatGPT 4.0 is less likely to get lost in its responses and remembers what was discussed earlier. On the other hand, in many cases cloud AIs will simply refuse to respond answering user questions due to software limitations.
Conclusions
It is great to see that AMD is confidently joining the AI race on all fronts: the release of new AMD Ryzen AI processors with NPUs and high-quality software adaptation for AMD Radeon 7000 graphics cards.
It is worth praising the software developers, as they managed to make Radeon graphics accelerators competitive with Nvidia in this area in such a short period of time. It is not only the fact that neural networks work, but also the availability of one-click installers that are suitable for users of any level of knowledge, that is important.
If you have an extra 100-200 GB of free space, a desire to experiment with different neural models, and modern AMD hardware, you should definitely try Amuse and LM Studio. Not so long ago, such privileges on Windows OS were available only to owners of green graphics cards.
A chatbot that swears and speaks without censorship and tolerance is at least interesting to talk to, and at most it will become an assistant in creative tasks. The same can be said about Amuse, which will definitely come in handy for artists, designers, and video game developers.
Spelling error report
The following text will be sent to our editors: