home.social

#dreambooth — Public Fediverse posts

Live and recent posts from across the Fediverse tagged #dreambooth, aggregated by home.social.

  1. CW: Stable Diffusion lawsuit

    Ok, I'm not sure if the plaintiffs involved actually know #StableDiffusion (SD) good enough to do that but unfortunately it could be relatively easy to fake evidence of SD straight up copying stuff by f.e. using #Dreambooth, #LoRa, #TextualInversion, #Hypernetworks, etc. to intentionally create a checkpoint/embedding/hypernetwork overfitted on a human-made piece of art and as judges are usually not very techsavy it's unfortunately possible they could be fooled by it.

  2. CW: Lensa, Stable Diffusion

    #Lensa ist also eine kostenpflichtige auf #StableDiffusion (SD) aufbauende App, die kann was normales SD mit #Dreambooth kann aber in schlechter und sogar teurer als z.B. Dreambooth über #HuggingFace (Ok, Lensa ist zugegebenermaßen jedoch anscheinend zugänglicher als normales SD mit Dreambooth, dennoch würde ich v.a. aus Datenschutz- und Flexibilitätsgründen von Lensa abraten).

    c't 3003 | Besser als Lensa! | Tutorial Stable Diffusion & DreamBooth youtube.com/watch?v=udvynl4Ycs

  3. Lensa AI app causes a stir with sexy “Magic Avatar” images no one wanted - Enlarge / A selection of male and female "Magic Avatars" generated by t... - arstechnica.com/?p=1904087 #machinelearning #stablediffusion #imagesynthesis #dreambooth #prismalabs #deepfakes #lensaai #biz&it #sexism #apps #ai

  4. Increasingly thinking that once I finish my current training project, maybe I should train ... either a or ... to specifically generate fellas. ;)

  5. Finally got a working, at long last... it literally took downgrading libraries and hacking the Dreambooth code itself, but it's running!

    First training samples are in. It'll be a long time before it's done, but I'm loving the start! So much better than what I was getting from .

  6. If anyone here has actually gotten to run - either locally on <12GB, or on a - please let me know. I've wasted too much of my life already on this :Þ

    Collab:

    GPU available: True, used: True
    ...
    CUDA_VISIBLE_DEVICES: [0]

    Local:

    RuntimeError: CUDA error: CUBLAS_STATUS_EXECUTION_FAILED when calling `cublasLtMatmul

  7. Testing two patches to for training - one to display multiple seeds at once during training, the other for automatic learning rate control. I'm trying a really difficult training task with ~1,5k images that needs high fidelity, so this would be better done with , but I can't get it to work on my RTX 3060. :Þ
    So I'm using a big hypernetwork. 50k steps in and there's still a long way to go.