IN NO EVENT SHALL BLACK FOREST LABS, INC. BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.

HyperFlux Diversity

I see ethnic diversity as an important part of image realism, and so I am delighted by the performance of dAIversity Flux (https://civitai.com/models/711900?modelVersionId=796248).

For most Flux models, I typically run dAIversity Flux with the AntiBlur LoRA (https://civitai.com/models/675581/anti-blur-flux-lora) to extend the depth of field. I am also captivated by the HyperFlux LoRA, as seen in the HyperFlux base model series (https://civitai.com/models/705444?modelVersionId=789074).

I had a performance issue with running the dAIversity Flux safetensors file with RuinedFooocus, which is optimized for GGUF use. So this factor and the desire to optimize the use of the two mentioned LoRAs with dAIversity Flux led to me creating a new merge in GGUF format.

HyperFlux Diversity is a blend of dAIversity Flux with the HyperFlux LoRA at strength 0.12 and AntiBlur at strength 1.0.

Please respect the licenses of each of the creators whose dedicated work created these resources. For example, do not use HyperFlux Diversity on any service that monetizes image creation, such as for online image generation. Also, do not sell or license this model for a fee or something else of value.

Usage:

My preferences for HyperFlux Diversity are a CFG of 3.5, but going up to 6.5 for special effects. I set sampling steps to 20 as I do for all 8-step Flux1 Dev based models, and use the euler sampler with the simple scheduler.

By default, HyperFlux Diversity creates images with a deep depth of field. Put "bokeh" in the positive prompt to create a shallow depth of field.

If this model creates NSFW images it is not intentional.

Credits:

dAIversity Flux: https://civitai.com/models/711900?modelVersionId=796248

AntiBlur LoRA: https://civitai.com/models/675581/anti-blur-flux-lora

HyperFlux 8 LoRA: https://huggingface.co/ByteDance/Hyper-SD

This project was created using this wonderful resource:
https://civitai.com/articles/8322/merge-a-lora-into-flux-for-better-speed-and-quantize-it

About the Images:

The first image in each GGUF version showcases a test generation at that particular quantization. Toggle through each version to compare how that image looks at that GGUF level.

The Q8 version provides the most detail, followed by the Q6. Both require 12 GB of VRAM. Q5 runs in 10GB of VRAM and it is almost as good. Q4 through Q2 have progressively less detail but occupy less space and will therefore load somewhat faster. It is a common misunderstanding that more highly quantized models occupy less VRAM but tests show that is not the case: Q4 through Q2 all use about the same VRAM as Q5, 10GB.

The rest of the featured images are samples from the ten standard tests applied to HyperFlux Diversity at the Pure Fooocus Base Model Reviews Guide:
https://www.facebook.com/groups/fooocus/learning_content/?filter=519238867322550&post=753347100153801
(Facebook links are notoriously unreliable, you may need to scroll up or down to access this particular guide)