***Work in Progress***
Hyper Flux Dedistilled models combine the Flux-dev-de-distill checkpoint by nyanko7 with the Hyper-SD LoRA by ByteDance into single-file models in q4_KM GGUF format for easy use on mid-tier hardware.
The Hyper 16 Steps version provides the same generation time as Flux-dev while producing better results. It approaches Flux Pro in terms of quality and prompt understanding. The Hyper 8 Steps version is twice as fast and offers a balance between speed and quality.
Various weights for the Hyper-SD LoRA were tested, and these models were created using the ones that provided the best results. This ensures that the models maintain consistent performance and converge as expected with the recommended step values, even after quality loss from quantization.
Versions and Recommended Settings:
Hyper 8 Steps:
CFG: 1 to 2 (not distilled guidance)
Steps: 8 to 12 (more is better, best at 12 steps)
Sampler: DPM-2 (for high detail) or DPM++ 2M (for faster generation)
Scheduler: Beta
Best for quick generations while keeping image quality above the standard Flux-dev level.
Hyper 16 Steps:
CFG: 2 to 3
Steps: 16 to 24 (more is better, but 20 gives great results)
Sampler: DEIS (for stylized, artistic outputs) or DPM++ 2M (for realistic raw images)
Scheduler: Beta
Ideal for refined, detailed images with excellent prompt understanding, similar to Flux Pro.
This is a pruned model, so you will need to download the following :
• Text Decoders: t5_xxl and clip_l from Flux Text Encoders (the FP8 version is recommended for mid to low tier pc).
• VAE: ae.safetensors from Black Forest Labs VAE.
I also recommend trying the following clip_l model for better text and details :
• CLIP Text Encoder (alternative option): CLIP-GmP-ViT-L-14 (HiT GmP TE).
No need to use Flux guidance—this model uses real CFG, just like any Stable Diffusion model, making it possible to use negative prompting effectively. For Hyper-16, use the same CFG value you would normally use for distilled guidance in Flux-dev (e.g., if you usually use 2.5, stick with that). For Hyper-8, I recommend using half of that value.
Future Versions:
This release is in q4_KM format. Other formats, such as FP8, q5_KM or q8_0, may be made available upon request for users with different hardware needs.
This project builds upon the incredible work of the original authors:
Flux-dev-de-distill by nyanko7: An advanced checkpoint implementing true classifier-free guidance.
Hyper-SD by ByteDance: A powerful LoRA solution enabling faster generation time.
The creation of these models is solely attributed to their respective authors. I take no credit for their work or development. The Hyper Flux Dedistilled models simply integrate and format these into a combined, accessible solution.
Both versions havw been tested to run smoothly on mid-tier GPUs, including a RTX 3070 and a RTX 2060 Super with 8GB VRAM, with excellent performance.
Description
Ideal for refined, detailed images with excellent prompt understanding, similar to Flux Pro.
Recommended Settings for Hyper 16 Steps:
CFG: 2 to 3 (real CFG)
Steps: 16 to 24 (more is better, but 20 gives great results)
Sampler: DPM-2 (best quality), DEIS (for stylized outputs) or DPM++ 2M (for realistic raw images)
Scheduler: Beta
FAQ
Comments (17)
gguf format? is that right?
yes
@C47HERINE Would it be possible for you to publish fp16 pruned versions as well? Sadly my software can't use GGUF format. And thank you for making this. I've been looking for something like Flux Fusion for dedistilled.
@anonborrower Sure! which version?
I can try both if you’d like. But, uploading on CivitAI takes a while and fails most of the time… would fp8 be fine ?
@C47HERINE I mention fp16 since bigger works better for me, but honestly I'd be happy with anything at all. Even with fp8, I could run tests to see how well this compares with my other workflows. There is no rush. I am very grateful that you are doing this for me.
@C47HERINE I'd be happy if there was an FP8 version which is not GGUF (Forge don't work with it). My GPU only have 8 GB memory, but FP8 is working well with it.
@anonborrower It's been updated with the 8 and 16 steps fp8 file! =D
@wregiszter632 There you go! I've up the safetensors for the two versions
The model works fairly fast on my 4070TI-Super and delivers an impressive realistic feel that I haven't felt with almost any flux dev (not dedistilled) checkpoint. Now if only this could be an uncensored checkpoint, it's frankly annoying how the base flux model insists on sfw only.
Thank you for the feedback! This model was more of a proof of concept to see if I could successfully tune a de-distilled version of Flux-dev. Now I’m working on a model that will take things to the next level. My goal is to create a version that can produce photographic, NSFW , and artistic results while maintaining a balance between speed, prompt understanding, and quality.
Working with de-distilled models requires some relearning. But, once you get the hang of it, the flexibility and potential are well worth it.
Thanks for your work! It's a pleasure to finally see some Lora actually working well. Anything I can do to help? I'd be happy to test and give feedback for instance. Hopefully if you manage to get a NSFW version it will allow for some true natural body, I am so fed up of the dozens upon dozen of flux versions with only ultra-shaved women. I don't get this! Woman's bodies are naturally beautiful - even older woman.
@sparkleboy599 if you are still interested I've made a brand new model of my own. The goal was to achieve this and much, much more! The dedistilled version is on its way, but as of now the Dev version is already available : https://civitai.com/models/1143800/pandora-flux
@C47HERINE thank you, i am going to check it out! :)
Never use a dedistilled before,is it normal that it takes double / triple of time to generate an image compared to normal flux?
Yes and the reason is that it requires more computational power. Compared to Flux-dev, when you use CFG, you’re basically telling the model how strongly to follow your prompt. A low value means the model takes some creative freedom, while a high value makes it stick closely to your instructions. This is classifier-free guidance (CFG), which balances what the model sees and what it should avoid.
The simplified Flux guidance works differently, it’s like a shortcut. It only focuses on basic clues, making it faster but less precise. Until now, this was Flux’s biggest limitation compared to other models like Stable Diffusion.
With dedistilled Flux, you get full CFG, resulting in better quality, more control, improved understanding, and negative prompting. All of this comes at the cost of longer generation times and more steps to achieve good results. The Hyper versions fix this by lowering the number of steps at a slight cost in quality. From my testing, Hyper-16 is about as fast as regular Flux-dev but delivers far better results.
Basically Dedistilled is OP compared to Flux-Dev 😆
@C47HERINE Thank you so much for the information! and thumbs up for your good work!!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.