Hello Everyone, Please read this before you use the Lora V4-Soap-Testing: I don't know if its good so I marked it as "testing". You can try if you want or you can stick to v3 or v2. If you want to use it, I STRONGLY recommend generating at 1344x1728 / 1440x1800 (gives the best results) or 1728x1344 Distilled CFG Scale: 3.5 Sampling method and Schedule type: DEIS with DDIM Steps: >=20 (I just use 20) Resolution: 1344x1728 / 1440x1800 (gives the best results) or 1728x1344. If its slow for you, generate at 896x1152 (gives the best results) or 1152x896 and use hires fix Lora Weight: 0.35 is good IMO. You can go to 0.4 or 0.5 also if you are not happy with the result. >0.6 is bad I think. You can get away with 0.6 if you generate at 2048x1536 maybe Positive Prompt: No need to use the prompt format of v2 and v3 (like "Amateur photography of" and end with "on flickr in 2007, 2005 blog, 2007 blog"). See the trigger word on the right side of this page (well, its not a trigger word but its how I am testing it and it worked well for me). Still, a detailed prompt always works best. See my example images to understand how I prompt it This is my first attempt at making a SOAP lora. If you test it, Please share your feedback. Thanks Recommended Settings (v3 and v2): Distilled CFG Scale: >= 3.5 (I use 4) Sampling method and Schedule type: Heun with BETA or DEIS with DDIM (Keep in mind that Heun doubles the generation time but it gives the best quality IMO - But you can use other sampling method as well. I just listed the one's I normally use) Steps: >=20 (I just use 20) Resolution: Flux is able to handle any resolution. I personally use 896x1152 or 1152x896 or 2048x1536 or 1536x2048 (All the training dataset images are 2048x1536 or 1536x2048 - with bucketing and training at 1024 resolution, it comes down to 832x1088 and 1088x832) Checkpoint: flux1-dev-Q8_0.gguf (Will work even better with the non-quantized original Flux Dev model but It is super slow on my machine so I use the Q8 one. I never tested this Lora with FP8 or Q4 or other quants) Text Encoder: clip_l and t5xxl_fp8_e4m3fn.safetensors (Will probably work even better if you use the full fp16 t5xxl but I don't use it because its slow on my machine) Lora Weight: Its mostly personal preference. I recommend that you do xyz plot and see which weight works best with your prompt (I test the lora at 0.8 weight) CFG: 1 Positive Prompt: If other prompts work for you with this Lora, just use that. I am just highlighting how I test the the Lora. I have seen several images here and in Reddit where people use different kind of prompts to get that "bad quality" look this Lora offers :D Always start with "Amateur photography of" and end with "on flickr in 2007, 2005 blog, 2007 blog" (I will probably remove this in v4 - its still WIP) 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 Negative Prompt: No (You can use it if you don't mind the generation time and add the words "blur", "blurry", "bokeh", "depth of field" to make the subjects sharper if they are blurry when you use this lora) How was the Dataset Captioned?: I captioned the training dataset using GPT4o. Starting from v2, every training image have >200 word captions. So if you want the best results with this Lora, you should describe the scene as detailed as possible Make sure you click on the "About this version" on the right hand side on this page to see the GPT4o prompt I used to caption the training dataset Shuffle captions option was not used when I trained the Lora Which version should I use?: I recommend using V3 but it can be tricky if you do not prompt it the way it likes. If you want some leverage, you can stick to V2 for now. V1 is not recommended Why am I getting blurry pictures?: If your whole picture is blurry (and if its not the look you are going for hahaha), its not an issue with this Lora. Flux base behaves like this sometimes. If that happens, change the sampler/scheduler/resolution and reroll the generation If the foreground is a bit blurry and the background is sharp, its related to this Lora. This Lora is not a SOAP (Shot On A Phone) type of lora. Its trained on some crappy 2007 photos so foreground blur happens with this Lora. To mitigate it, I suggest that you try to generate at higher resolutions or use hires fix If the foreground is sharp and the background is blurry, you should probably remove any "beauty" or any focus-related terms from your prompt Why am I not getting the Amateur quality this Lora promises?: It happens to me as well. There are various reasons like the prompting is not like how this lora likes it, the prompt is good but none of the words I used in training are in the prompt or the weight of the lora is too low. I can recommend a few things Try to increase the weight of the Lora until you get what you are looking for. If what you want is at weight 1 and the image have artifacts, use hires fix or some other upscale methods Use the GPT4o prompt I provided to get the prompt. After you send my prompt to chatGPT, you can just ask it to give you a prompt like "Give me a prompt of a woman riding a buffalo in las vegas" Just experiment haha. I am new also to this and I am still learning If you can donate Buzz, its appreciated. Flux lora training is not cheap in civitai haha. If you cannot donate any Buzz and you like this Lora, please like it If you do not like it, please leave a comment explaining where its struggling and I will try to fix them in the next version. Thanks https://civitai.com/models/652699?modelVersionId=828456 ---- (Read the model description) I don't know if its good so I marked it as "testing". You can try if you want or you can stick to v3 or v2. If you want to use it, I STRONGLY recommend generating at 1344x1728 (gives the best results) or 1728x1344. If its slow for you, generate at 896x1152 (gives the best results) or 1152x896 and use hires fix