Showcase Images were created with Fluxmania or Flux1.dev fp8 in PNG and include metadata, prompt & workflow. All licensing conditions for the Flux Dev model apply. You can find the details here.
Description
Model trained on Civitai
Recommended resources : Fluxmania III
Recommended settings : dpmpp_2m / sgm_uniform / 25 - 30 steps / flux guidance : 3.5
Weighting : 0.8 to 1.0
FAQ
Comments (12)
Looks good. No wipeouts with this LoRA. Can you recall if any of your training data had images of surfers inside the barrel of the wave?
Refer to the link for example:.
Thanks for the feedback, there was only one photo (on 69 photos) in my dataset with a surfer in the tube riding
@Adel_AI ok without having to trial and error too much to see if that 1 image can be replicated/generated, do you think it might work if we prompt for a scene like this (inside the tube/barrel) but my question is if we do prompt, do we need to use similar prompts/captions that you used in your training (eg. tube) or if we prompt using genuine surfing terminology (eg. barrel) instead of tube, the lora will understand us better?
@xlr8td Really, I haven't had time to properly test this model. I think that the 2 approaches should be tested but I doubt that the terms specific to surfing will be recognized
@Adel_AI OK I understand, some authors do use official terminology while other authors don't have the time to provide relevant terminology related to the topic trained. I didnt want to waste images trying to get the lora to generate prompts it didnt understand. My best approach will be to guide it step by step as if it is an infant/child and describe the shape of wave etc and so on. I am aware you are constantly releasing loras but if you where interested in using my images of the fishes in flower gardens to train a lora on that data, let me know, I could do it myself but I will probably not get good training results like you or others would. Unless you have a workflow you can provide for me to follow for the training, let me know.
@xlr8td all my lora are trained on Civitai, I do not have enough memory to do it locally.
There is a trick to reproducing particular compositions, without having to heavily detail the prompt. Simply do I2I with a desired image, use it as a noise source and leave the denoising at 1. This method is more efficient than T2I. The result is better if it is combined with caption with Florence2 to provide a simple prompt, not too detailed.
This technique leaves more creativity to the model than using controlnet
Thanks for the info but I don't have a local setup myself either, when I was saying I could train it myself I was refering to using this online service to train but since you have had a lot of luck with training good loras I thought I would offer to you if you wanted to train it and we split the buzz earned (I think the default is 50% each for collaboration work). Let me know what you think. As for the information you provided I am not sure if it refers to training lora or refers to generating the surfing images that I want to generate? Also if you do consider doing the training of the fishes, please take screen captures of all the settings you used before training etc so if I do any training myself I just follow your settings/tutorial and I dont ever have to ask you again in the future if you want to train something. Also I may have to generate some more fish varieties so you have enough images to train.
@xlr8td you have certainly noticed that my LoRA publication frequency has dropped significantly because I am overwhelmed with work. I have almost fifty LoRA projects that I want to finalize but I can no longer find the time. If you would like to have my advice throughout the development of your LoRA project, I have no problem, and it is without compensation in Buzz.
@Adel_AI ok thank you, let me know when you are available in the month of March, plenty of time until then and I will get more images generated for training lora for Flux. Is 20 images enough or you recommend at least 30 images? That info you provided previously, was it for training or was it a trick for getting an image generated without using controlnet (eg. provide an image of a sufer inside the tube of wave and select image to image?)
@xlr8td it is preferable to continue this discussion in a private message
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.