.jpeg)
This is a workflow that I put together in order to use Flux Dedistilled, Dev2Pro, and all of their variants. This workflow will work with GGUF versions of any and all safetensor versions of flux dedistilled variants, although I've only tested the gguf8 and the fp16 versions of the models. I use the "ViT-L-14-TEXT-detail-improved-hiT-GmP-TE-only-HF" text encoder as it yields vastly superior results in my humble opinion, however, use whichever encoder you would like. Also, the other clip model I use is the "t5xxl_fp16" encoder, and I DO believe it will work with the lower quantized versions, I have tested it with GGUF8 versions of the models, so I can confirm it works on the gguf. IN this workflow I use the fp16 version of the flux nude men and women model which is a merge of flux dev2pro and flux dedistilled, if I'm not mistaken. In my humble opinion this is the best of both worlds, you have reasonably good NSFW capabilities without needing loras, and the "clothed" content yields very good, natural, unique results. This workflow contains the "FluxModelingSampler" which I HIGHLY recommend if you are using flux, there are a few explanation videos and tutorials on youtube on how to use this feature to literally unlock flux and get it to do incredible prompts with practically unmatched detail. Also, the latent image section consists of 3 nodes, the width, the height, and the actual latent image generator. This gives you the same control over your height and width but it also helps insure that the output sizes are quantized to a more "computer friendly" math, I.E, if you wanted 1024 by 1820, it automatically moves it up 4 pixels to 1824. And yes, this does seem to help models converge better and render more consistent output over various seeds and lora weights. Not sure why, and this is just my observation. This workflow is based directly on two different workflows that I found in the comment section, I basically picked them both apart, found what works best, then slammed the two together. The two original contributors are: Vinque and Afroman4peace, who's workflow was posted in the flux dev dedistilled main page by hannathorn. Thank you for your contributions and link posts.