[STOP! Give GGUF a try! You'll LOVE it!!! (more below)]
A merge of FluxUnchained and FastFlux. As a result, it can now generate artistic NSFW images in 4-8 steps.
You can also combine it with other LoRAs to get the effect you want.
Which model should I download?
T5-F16: T5 in full precision is included. However, requires more VRAM.
T5-F8: T5 in fp8 is included, requires less VRAM for T5.
UNet_only is Recommended. Download and use your own separate T5xxl, CLIP and VAE. See https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/1050 to learn more about where to download the VAE, clip_l and t5xxl models.
GGUF is the Future
TRY THE GGUF version at https://civarchive.com/models/662112 . It is faster, better, and more convenient.
GGUF Q8_0 is a more precise format than fp8 and also runs faster (at least on my 1080ti)
You really shouldn't be using a flat safetensor that includes T5 if you can. Every new checkpoint adds a new copy of T5 (additional 10GB/5GB) that you don't really need.
That being said, some UI/softwares still don't support GGUF / separate files. That is why I uploaded these safetensors. But I feel guilty contributing to this anti-pattern.
Description
T5(bf16), CLIP and VAE included.
FAQ
Comments (14)
is possibility for flux dev and Q8?
Ofc. https://civitai.com/models/662112?modelVersionId=748232 . But since this is a dev model you'll need ~20 steps for good results.
@nakif0968 20 steps is not a problem - quality is more important ;)
Dd you consider also make Q4k_m version as has much better quality than "old" Q4?
Q4k_m has quality as "old" Q5.1.
Thanks
thank you
@mirek190 I've been using python to convert the safetensors into gguf. I don't think the gguf-py library supports advanced quantization like Q4_K_M/S (it can only dequantize them). If this is added I can proceed. Or you'd have to ask someone smarter than me... :P
@nakif0968 Here you go :0
You can ask him. O
Or I can do that tomorrow.
Keeps coming up on the list as Updated, what are the updates?
Civit AI bug maybe
Words for some body parts were censored in base Flux - do you know whether this has been retrained to include recognised trigger tokens for these?
I can't tell. It does seem to respond to keywords like nipples and vagina.
Whatever you are doing training Flux is good. It hasn't lost prompt adherence as far as I can tell.
请问，你是用什么软件制作flux模型的？我不想只限于lora制作！
请问，你是用什么软件制作flux模型的？我不想只限于lora制作！
FLUX unchained is a merge of multiple loras, rather than a fine-tune. If you want to see how a real fine-tune is made, see PixelWave or SapienF
There is this new quantization technique called svdquant nunchaku, which is smaller is size and very fast, about 3x times faster. Any plans to release your unchained model in this svdquant format.