This set of models for SDXL and Flux are various merges that attempt to address different use-cases like 2d logos, weird-odd styles, or realistic photos. The focus has been to maintain quality and get clean/crisp output that can be upscaled for very high quality.
The 1.0 (Base) model is mostly tailored to 2D content like logos, stickers, anime, etc...
The Weird model is focused on odd, weird, and darker themed content.
The Realistic model is focused on photos, photography, and selfie type imagery.
The Flux model is solid in all aspects.
Description
This model works just fine at 10 Steps. In fact I never use more for any of the images I generate with it.
This is a merged model from many different Flux models each merged at different percentages.
It is a really good all around model.
Like with most Flux models be sure to use the standard Text Encoders and VAE.
FAQ
Comments (10)
Is the Flux version the fp8 version?
works very long - compared to other Flux models.
It's a Pruned version of several merged FP8 versions. I average around 20 seconds for an individual image and about 1 minute for 4 images at 1152x896, 10 Steps. This is in Forge SD.
This looks like NF4 version
flux version upload the wrong file? have the same hash as https://civitai.com/models/630820?modelVersionId=983597
I don't think so. I'll check and confirm. This is a merge of like 5 different models at various percentages.
I believe it did not update the Hash value because there is an issue in my Forge instance where you have to disable the metadata updates or it won't merge models.
that flux file at 6gb seems like the wrong one, is it a quant?
It's a merge of various formats FP8, NF4, etc.., but a pruned version.
Will work with LORAS?
@zerocool22 Yes