Most likely the last model for Flux for the time being, unless something interesting pops up. Reasons are manifold, but the main reason is i miss the randomness of SDXL and Pony models which was the most fun. Pretty chill listening to music and taking the lyrics appart in between to see what happens,... can't really do that with Flux. Also, the resources required along with time to generate a picture are just to much, especially at the moment. Add a few LoRA's, a higher CFG and the RAM usage + generation time goes through the roof. In the same time it takes for one picture i could have made a batch of 10 with High-Res Fix in SDXL. Will still use it here and there, but not as the main focus.Curiosity and banging my head against the wall was the main driving force to even start with Flux. There are a lot of benefits with Flux, but after looking into it a bit more, it's a fight i can't win when it comes to the way i want it to behave or what i expect/hope it would do. Unflexible, defaults to certain images and styles even with LoRA and random seed, concept bleeding ( which is in some cases not a bad thing ), some plugins don't work ( either not supported or Gradio version in Forge ), limited knowledge about camera angles, some images have a cartoony fake look to it even realistic pictures, limited noise, grain, haze, blabla etc.What it excels at is hower, very good for styles, clean images, turns even bad training images into something useful, good anatomy for the most part, color and contrast, coherence and text encoder etc.This model wasn't trained on the default Flux model, but this one .<a target="_blank" rel="ugc" href="https://civarchive.com/models/843551/fluxdev-dedistilled?modelVersionId=943787">https://civarchive.com/models/843551/fluxdev-dedistilled?modelVersionId=943787</a>. Does it do something different? Dunno. I would've done a side by side comparison training the same LoRA with both versions, but didn't really bother in the end. However, that being said, i would kinda discourage people that are interest to use this LoRA with the on-site generator, unless you wanna burn Buzz on something that might not work.There are quite a few LoRA's like this atm ( well, Halloween ), which also do a better job i guess. Even though it doesn't really show it, because all pictures posted are pretty similar, i used several images with different styles and theme, some don't even have anything to do with horror. Just used the current Halloween stuff as the main focus and to test some theories or hunches for whatever it's worth.Good LuckPS: Some images before in my gallery where created with basically the same LoRA, just different settings in Kohya and step count.

Flux.1 D - Creepy Stuff

- Flux1 De-Distilled - temp_haunting_cre_1.0.safetensors

v1.0

Most likely the last model for Flux for the time being, unless something interesting pops up. Reasons are manifold, but the main reason is i miss the randomness of SDXL and Pony models which was the most fun. Pretty chill listening to music and taking the lyrics appart in between to see what happens,... can't really do that with Flux. Also, the resources required along with time to generate a picture are just to much, especially at the moment. Add a few LoRA's, a higher CFG and the RAM usage + generation time goes through the roof. In the same time it takes for one picture i could have made a batch of 10 with High-Res Fix in SDXL. Will still use it here and there, but not as the main focus.

Most likely the last model for Flux for the time being, unless something interesting pops up. Reasons are manifold, but the main reason is i miss the randomness of SDXL and Pony models which was the most fun. Pretty chill listening to music and taking the lyrics appart in between to see what happens,... can't really do that with Flux. Also, the resources required along with time to generate a picture are just to much, especially at the moment. Add a few LoRA's, a higher CFG and the RAM usage + generation time goes through the roof. In the same time it takes for one picture i could have made a batch of 10 with High-Res Fix in SDXL. Will still use it here and there, but not as the main focus.Curiosity and banging my head against the wall was the main driving force to even start with Flux. There are a lot of benefits with Flux, but after looking into it a bit more, it's a fight i can't win when it comes to the way i want it to behave or what i expect/hope it would do. Unflexible, defaults to certain images and styles even with LoRA and random seed, concept bleeding ( which is in some cases not a bad thing ), some plugins don't work ( either not supported or Gradio version in Forge ), limited knowledge about camera angles, some images have a cartoony fake look to it even realistic pictures, limited noise, grain, haze, blabla etc.What it excels at is hower, very good for styles, clean images, turns even bad training images into something useful, good anatomy for the most part, color and contrast, coherence and text encoder etc.This model wasn't trained on the default Flux model, but this one .<a target="_blank" rel="ugc" href="https://civitai.com/models/843551/fluxdev-dedistilled?modelVersionId=943787">https://civitai.com/models/843551/fluxdev-dedistilled?modelVersionId=943787</a>. Does it do something different? Dunno. I would've done a side by side comparison training the same LoRA with both versions, but didn't really bother in the end. However, that being said, i would kinda discourage people that are interest to use this LoRA with the on-site generator, unless you wanna burn Buzz on something that might not work.There are quite a few LoRA's like this atm ( well, Halloween ), which also do a better job i guess. Even though it doesn't really show it, because all pictures posted are pretty similar, i used several images with different styles and theme, some don't even have anything to do with horror. Just used the current Halloween stuff as the main focus and to test some theories or hunches for whatever it's worth.Good LuckPS: Some images before in my gallery where created with basically the same LoRA, just different settings in Kohya and step count.

Files with SHA256: 4423106ab06b0e3457bc306762aa501ee748bcff68cbe92a8c05a356797a8f98