Introduction
If you want to know who the most popular and promising open source large model in the AI painting world in 2024 is, it has to be FLUX!
First, let’s take a look at the rendering results in SD1.5, SDXL, SD3 and Flux using the same prompt word. It is obvious that Flux has better results.
Prompt word:
Picture quality: highest quality, true photo style,
People: a young woman wearing a bright blue dress
Details: detailed eyes, delicate face, pore details, detailed eyes, pupils, irises, delicate lips, beautiful face, wearing a necklace, adding a touch of elegance to the scene
Environment: standing in a gathering surrounded by a crowd of people, surrounded by bright lights
Picture size: 1280×768
Sampling method: euler + simple
Sampling steps: 20
Guide coefficient: 7 CLIP SKIP: 2
The reason why Flux’s effect is so outstanding is inseparable from its training parameters of up to 12 billion. A single model can reach a terrifying 22GB (the computer hard disk is not enough).
Advantages of Flux
- The picture effect is good. Just by relying on the prompt words, it can generate lifelike pictures, which is definitely the strongest among all open source large models. The picture quality, details and texture are all amazing.
- It solves the problem of AI painters. Whether it is anime style or real-life style, the details of the hands are very well portrayed.
- The function of outputting text is unexpected. As long as your prompt words are detailed enough, you can express any text in any scene. Although it does not currently support Chinese, it will surely do so in the near future.
- It has large training parameters and many styles.
In general, you don’t need to download many large models with different styles. You only need a base model to easily control creative painting in different styles.
- Flux drawing has great flexibility.
Flux drawing greatly reduces the requirements for prompt words and differentiation.
For example, there is no longer a need to enter negative prompt words, and there is no need to carefully select different drawing sizes for different models.
Various models derived from Flux
Black Forest Official has released three Flux models: Pro, DEV, and Schnell
- the Pro version only supports API calls, and under normal circumstances, the best version we can use is the DVE version.
- the DEV version has the best FP16 effect, but if you have a graphics card below the 90 series, don’t even consider it. If you don’t have a top-of-the-line graphics card, you can consider the FP8 version, which only has half the capacity. The difference between the FP16 and FP8 rendering effects is not significant, and it will only be different in very detailed places, and you may not be able to see the difference if you don’t look closely.
- The Schnell version is equivalent to the accelerated version of the DVE version. It can reduce the number of sampling steps from 20 to 4-8, which greatly reduces the waiting time for rendering, but the rendering effect will be relatively worse. However, the rendering effect is still very good.
But even the fastest and smallest Schnell FP8 version, the video memory usage has reached more than 14GB.
- The faster GGUF version.
GGUF breaks Flux into small models of different sizes, with capacities increasing from Q 2 to Q 8. You can choose the appropriate version according to your graphics card.
- NF4 version. It integrates VAE and T5 text encoders, so it has a larger capacity of 11GB, and a graphics card with a minimum of 8GB video memory is required to use it.
Below is a comparison of the rendering effects of each version of Flux.
