Fizgig Model EQ – ACE-Step 1.5 XL Turbo
ACE-Step 1.5 XL Turbo's own attention weights as a rank-128 LoRA: the top singular directions of all 256 attention layers, by exact SVD, about 42% of their weight energy. Made for Fizgig's Repair Studio, where its block sliders become an EQ of the model itself: 1.0 is the stock model, 0 and 2 are −25% and +25% of each layer's main directions. Shape the model by ear with no training, then save the result as an ordinary LoRA that loads anywhere at strength 1.0.
Not meant to be loaded directly in ComfyUI. Use the LoRAs you save from Fizgig.
Base model: ACE-Step 1.5 (MIT). Made with Fizgig.