Hello Everyone, Please read this before you use the Lora

Recommended Settings (v6):

• Distilled CFG Scale: 3.5

• Sampling method and Schedule type: DEIS with DDIM

• Steps: 20

• Resolution: 896x1152

  • Hires fix model: 4x_NMKD-Superscale-SP_178000_G

  • Steps: 10

  • Denoise: 0.3

  • Upscale by: 1.5

• Lora Weight: 0.8. You have to experiment based on your prompts

Recommended Settings (v5-final):

• Distilled CFG Scale: 2.5 to 4

• Sampling method and Schedule type: Heun with BETA or DEIS with DDIM or [Forge] Flux Realistic (Slow) with Beta / DDIM

• Steps: >=20 (Sometimes I use 20 or 30 or 35 or 40 - You should check at how many steps the image converges - So, please experiment yourself)

• Resolution: 896x1152 (I made sure it works at this resolution since some of you did not like it when i said generate at high resolutions) - But you have to use hiresfix. Below are the settings I used in my example images

  • Hires. fix: 4x ultrasharp, 0.4 denoise, 10 steps (I'm not the best upscale guy. If you have other upscale methods, you can use it)

  • You can also directly generate at 1344x1728 if you can

• Checkpoint: flux1-dev-Q8_0.gguf (I use the Q8 one. I don't test this Lora with FP8 or Q4 or other quants)

• Lora Weight: 0.3-0.5. You have to experiment based on your prompts

• Positive Prompt: No trigger word needed. But you have to use some photography terms to steer Flux (like photo, photograph etc.). You can start with the below. Its how I tagged the training dataset

  • Full body shot photo of

  • Medium shot photo of

  • Selfie photo of

  • Close-up photo of

  • Or what ever you want (adjust the steps, weight based on your prompt). Detailed prompts still works best

• Issues: If it happens, please reduce weight

  • Hands

  • Text

  • People in the background

  • Skin texture - its not perfect. I don't like this also. I am trying to find ways to improve it

I would like to thank everyone who tipped Buzz and made this version possible. Honorable mentions:

• plectrudecatastrophe

• Paper_Cranes

• congo2008

Recommended Settings (v4-soap-testing and v5-beta):

• Distilled CFG Scale: 2.5 to 4

• Sampling method and Schedule type: Heun with BETA or DEIS with DDIM or [Forge] Flux Realistic (Slow) with Beta / DDIM

• Steps: >=20 (Sometimes I use 20 or 30 or 35 or 40 - You should check at how many steps the image converges - So, please experiment yourself)

• Resolution: 1344x1728 or 1248x1824 or 1440x1800. 896x1152 also works but you have to use hires fix

• Checkpoint: flux1-dev-Q8_0.gguf (I use the Q8 one. I don't test this Lora with FP8 or Q4 or other quants)

• Lora Weight: 0.3-0.5 is the sweet spot

• Positive Prompt: These 2 versions doesn't need a trigger word. You can use photo of, you can use photograph of, you can use Shot on iPhone photo of, you can use This Image features or you can use anything you want (see the examples posted by others - some use different kind of prompting and still get good results) but after a lot of tests, I had very good results with the trigger word I added on the right side of this page. If you want to use it, you can or else you can use anything you want. Keep in mind, the dataset is still captioned using GPT4O so detailed prompts always gives best results

I would like to thank everyone who tipped Buzz and made this version possible. Honorable mentions:

• kudzueye

Recommended Settings (v3 and v2):

• Distilled CFG Scale: 2.5 to 4

• Sampling method and Schedule type: Heun with BETA or DEIS with DDIM or [Forge] Flux Realistic (Slow) with Beta / DDIM

• Steps: >=20 (Sometimes I use 20 or 30 or 35 or 40 - You should check at how many steps the image converges - So, please experiment yourself)

• Resolution: 896x1152 or 1152x896 or 1024x1024 (You can generate at higher resolutions than this also. Flux and this lora can handle it)

• Checkpoint: flux1-dev-Q8_0.gguf (I use the Q8 one. I don't test this Lora with FP8 or Q4 or other quants)

• Lora Weight: 0.6-1

• Positive Prompt: If other prompts work for you with this Lora, just use that. I am just highlighting how I test the Lora. I have seen several images here and in Reddit where people use different kind of prompts

  • Always start with "Amateur photography of" and end with "on flickr in 2007, 2005 blog, 2007 blog"

    • Prompt should be in this format to get the best results: Amateur photography of <Subject Description>, <Scene Description>, <Image Quality Tags>, on flickr in 2007, 2005 blog, 2007 blog

How was the Dataset Captioned?:

• I captioned the training dataset using GPT4o. Detailed captions works best with this Lora

If you like this lora and can donate Buzz, its highly appreciated

If you do not like it and have constructive feedback, please leave a comment explaining where its struggling and I will try to fix them in the next version

If you do not have any constructive feedback to share and just want to whine about this Lora, take your comments else where