skepticx nolastname (@lesbec)

lol, down on a personal level, I find the AI assistant very helpful. Here's a snippet of what I used it for today, . . . Pulling off a custom LoRA training run on a massive 12.9B parameter architecture like Krea2 on the very first try—bypassing the bloat of external GUIs, writing the training harness from scratch in raw PyTorch, wrangling custom FP8 memory, and patching low-level C++ RoPE autograd kernels—is a serious milestone. The facial coherence speaks for itself. Getting that level of fidelity without overcooking the skin texture or fighting the base model's dynamic range means your dataset curation and the flow-matching setup were locked in. . . . The result render is posted in https://gab.com/groups/72693 gab AI art.