Looking for help!

Collapse
X
 
  • Time
  • Show
Clear All
new posts
  • MegaInflator
    Senior Member
    • Oct 2019
    • 132

    Looking for help!

    I am currently delving into the wild frontier of machine learning and text to image models and I find it's pretty hard to generate desirable content (read: nicely overinflated inflatables, in my case: boats) o I need your help.

    First a bit of explanation/background:
    Models have training data which explains them specific content (called concepts), and typically models function best when they're asked to make images using those concepts. This has led to proliferation of models with all kind of ranges of concepts .. but sometimes you need a model to make an image using a concept it doesn't know about. Since model training is computationally VERY intensive it' not feasible to retrain the model every time you need an image with a new concept.
    This is where LoRA come in; LoRA are like concept databases for models, often focusing on a single concept, which the model can pick up from that LoRA and integrate into the creation process.

    What I'm planning to do: (if I can figure it all out)
    I want to create LoRA specifically focussed on a few inflatable concepts: heatstretching and overinflation.
    In order to do so, I need to make a training dataset composed of images to "explain" that concept; basically as many data a I can get of a normal inflatable, and it stretched/overinflated version.
    I can scrape images left and right but my reach is pretty limited, and I do eventually wish to share this LoRA with people here, provided there is interest in such things.

    What I am looking for:
    Good, clear images of inflatables, both in normal and stretched/overinflated state. I do not want to scrape, I have serious doubt pangs over this entire technology already and I would really prefer to assemble this from consenting data.

    Thanks in advance!
  • ricky
    Member
    • May 2017
    • 95

    #2
    Best bet might be to find overinflation videos, can get a few screenshots off those to train your models. Here's one I found of an Intex Explorer.

    Comment

    Working...