If you like what I do, leave me a ❤️ to receive notifications of future updates and follow me here on Civitai. Remember that it is a hobby for me and it costs a lot of time and everything is shared entirely for free, I appreciate that you make a review and share your creations on this wall.
Project HaDeS - FLUX S - V1
From what I learned with Flux Dev and my Project Gaia, I have decided to invest some time in thoroughly studying the smallest of the Black Forest Labs models, Flux Schnell.
Project HaDeS is a fast model, but with more quality than the original, at least in the areas that have been trained. With each update, I will be retouching the original dataset, improving each layer of the model, except for the text, which works, in my opinion, more than well.
In this first version, I have touched the male models, black female models, various textures such as skin, stone, earth, water. I have incorporated some LoRa from my collection, such as more realistic water, meaningful tattoos and improved the definition in general.
Recommended Settings:
Resolution:
The SDXL ones work very well in my opinion, although in distant images some quality is lost if the image is not rescaled.
For those who don't know what resolutions I'm talking about, they are the following:
1:1 exact 1024 x 1024
3:2 exact 1254 x 836, rounded 1216 x 832
4:3 exact 1182 x 887, rounded 1152 x 896
16:9 exact 1365 x 768, rounded 1344 x 768
21:9 exact 1564 x 670, rounded 1536 x 640
For those of you with a good graphics card, I recommend the higher resolutions, which are the maximum that the flux1 models allow before losing quality and everything becoming blurry:
1:1 exact 1448 x 1448, rounded 1408 x 1408
3:2 exact 1773 x 1182, rounded 1728 x 1152
4:3 exact 1672 x 1254, rounded 1664 x 1216
16:9 exact 1936 x 1089, rounded 1920 x 1088
21:9 exact 2212 x 948, rounded 2176 x 960
Recommended steps:
For the vast majority of images, especially portraits and close-ups, 4 steps are more than enough. For very detailed images, increase the steps until you find the best one.
CFG and Distilled CFG:
1 and 2 respectively. Distilled can be raised up to 5, but 2 is a value that provides much more realism in my opinion.
Upscale method:
Ultimate SD Upscaler gives good results, I'm also testing Hires upscale with the 4xNomos8kDAT.pth model, but I can't recommend it yet, as I haven't tested it enough. But I'll leave it here in case anyone is interested.
Sampling method:
I've tried a few, but I still think Simple Euler is the best. I'll leave that up to you, though, as there aren't too many at the moment to spend hours trying.
I hope you like it. Post your images on the wall, give it all your likes and love, and see you next update!!
Description
Small improvements in quality and definition.
FAQ
Comments (10)
The model looks very interesting! Do you plan to release some other format, not NF4, for those who can not run the model of this format?
I can upload the Unet only if needed
@DeViLDoNia That would be great, thank you!
@Diana111 Unet version is available.
It took me forever to find references on how to convert to GGUF, but once I figured it out, it's pretty easy.
@moxie1776 I'm just making pictures! I don't know how to work with models other than that. It's interesting, but there's no point getting into it with 2080 8Gb.
@Diana111 have you tried a GGUF Q4? It's about the same size as an SDXL model. Here is one I posted, but there are several, meant particularly for low memory cards. If you have to go below a Q4 to a Q3 it probably wouldn't be worth it, but if you can run SDXL you should be able to run a GGUF Q4 Flux model.
https://civitai.com/models/781094?modelVersionId=944360
@moxie1776 My husband bought me 64 Gb RAM, and I can run FP8 and BF16 models in normal image size. So I don't really need the NF4 format, but the Project HaDeS model looked amazing. I've tried all Flux.Schnell models on CivitAI that don't prohibit commercial use, Project HaDeS is currently one of the best.
Unfortunately, Flux.Dev has an unclear license, and I use some of the images for commercial purposes.
@Diana111 NF4 is just a lower memory requirement, and would probably run better for you. GGUF takes a quality hit, but would probably run a lot better on 8GB of VRAM. Anyway, have fun, that's what it's all about.
@moxie1776 Thank you!