Warning: Although these quants work perfectly with ComfyUI - I couldn't get them to work with Forge UI yet. Let me know if this changes. The original non-k quants can be found HERE which are verified working with Forge UI.
[Note: Unzip the download to get the GGUF. Civit doesn't support it natively, hence this workaround]
These are the K(_M) quants for HyperFlux 8-steps. The K quants are slightly more precise and performant than non-K quants. HyperFlux is a merge of Flux.D with the 8-step HyperSD LoRA from ByteDance - turned into GGUF. As a result, you get an ultra-memory efficient and fast DEV (CFG sensitive) model that generates fully denoised images with just 8 steps while consuming ~6.2 GB VRAM (for the Q4_0 quant).
It can be used in ComfyUI with this custom node. But I couldn't get these to work with Forge UI. See https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/1050 for where to download the VAE, clip_l and t5xxl models.
Advantages Over FastFlux and Other Dev-Schnell Merges
Much better quality: you get much better quality and expressiveness at 8 steps compared to Schnell models like FastFlux
CFG/Guidance Sensitivity: Since this is a DEV model, unlike the Hybrid models, you get full (distilled) CFG sensitivity - i.e., you can control prompt sensitivity vs. creativity and softness vs. saturation.
Fully compatible with Dev LoRAs, better than the compatibility of Schnell models.
The only disadvantage: needs 8-step for best quality. But then, you'd probably try at least 8 steps for best results with Schnell anyway.
Which model should I download?
[Current situation: Using the updated Comfy UI (GGUF node) I can run Q6_K on my 11GB 1080ti.]
Download the one that fits in your VRAM. The additional inference cost is quite small if the model fits in the GPU. Size order is Q2 < Q3 < Q4 < Q5 < Q6. I wouldn't recommend Q2 and Q3 unless you absolutely cannot fit the model in memory.
All the license terms associated with Flux.1 Dev apply.
PS: Credit goes to ByteDance for the HyperSD Flux 8-steps LoRA which can be found at https://huggingface.co/ByteDance/Hyper-SD/tree/main
Description
FAQ
Comments (5)
I just finished testing this model at the Pure Fooocus Facebook group (https://www.facebook.com/groups/fooocus). It is a default Fooocus model in the SimpleSDXL branch of Fooocus.
HyperFlux 8 Q5_K-M produced excellent results overall, scoring 84%, and currently in second place. For best results I run it at 20 steps, the same as Flux1 Dev - and for me it makes Flux1 Dev and Flux1 Schnell completely obsolete.
I was extremely impressed at how well it performed in my Long Prompt test, achieving a perfect score that only one other model has achieved. And the quality of imagery in doing so was breathtaking.
My only criticisms are the lack of depth of field, a trait is shares with other Flux models. However HyperFlux 8 Q5_K-M responds very well to the https://civitai.com/models/675581/anti-blur-flux-lora which effectively negates this shortcoming.
There was also a surprisingly poor response to my melanin wildcard, with the model ignoring subtleties in skintone variations. However it is capable of producing many skintone variations if prompted for ethnicity.
I really like this model. But because of .zip it came without meta data. Do you know a workaround?
Is there any difference in the drawing quality between Q4_K_S and Q4_K_M, and Q4_1 and Q4-0? Is it just about file size and VRAM size? thank you
Bigger models usually give better quality.
Yo! I had to leave a comment for this one reason alone... The "GPUS COST TOO MUCH" image is Freaking Hilarious!!! LMAO! Genius art! I can only hope to be as creative in near or distant future.