Hi there,
this is my first FLUX finetuning test. I startet low and just reused my already existent landscape painting dataset and took only the best ones that define the essence of the Hudson River School of art. I trained a LORA and you can use the output however you will but if possible I would like to be informed, just out of curiosity what you guys do with that.
Tag: HRScontrast
You may want to add a "landscape" afterwards, with a space in between. I found that a bit overtraining it and then using a lower strength (like 0.5) works best. I am very much open for feedback and in a not too far distant future there will be a full finetune based on my dataset.
Description
trained on 40 paintings, 20 steps each, 3 epochs. Use <1 strength
FAQ
Comments (3)
Great job!
I would LOVE to see the training data/captions so I can understand how you captured this so well
Do you know a good way of distributing the training data?
All the picutres were first captured with COGVLM and afterwards I cleaned the captions heavily, so that now they are all basically manually captioned. It is not fully natural language but rather a mixture of tags (like, "traditional oil painting") and short descriptive half sentences ("on the left are conifer trees"). The captions are quite short relative to what we normally do with flux and I may have to revisit the captions. However since the full dataset is over 600 pictures big that I all manually captioned I have to think about a good way of doing that.
I also think about an upscaling lora or finetune with the pictures zoomed in so that the model gets an understanding about the details it is supposed ot add.
@Kornrat Civitai allows you to post the training data, I think. I've seen it here. Other than that, drop box, or a Gdrive link?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.