<h1 id="krea2-finalcut">Krea2 FinalCut</h1><h2 id="v3-major-update">V3 Major Update</h2>Both Universal and Comfy Quant versions available, please read <a target="_blank" rel="ugc" href="https://civarchive.com/articles/35231/how-much-quality-is-lost-fp16-vs-fp8-vs-int8">this article</a> if you have any question on what format to use. (Mac Users should likely use FP16)<h3 id="nvfp4">NVFP4</h3>If your Blackwell is showing <code>autocast BF16</code>, don't expect the major NVFP4 generation-speed improvement.<blockquote>"NVFP4 makes the model smaller and can make loading/switching LoRAs faster. On Blackwell GPUs, you can also get a real generation-speed improvement—but only if the model is actually being calculated in NVFP4. If your software automatically converts the model to BF16, you aren't getting that special Blackwell NVFP4 speed boost. You're mainly benefiting from the smaller model and faster loading. Older GPUs can get those loading benefits too.The trade-off is that NVFP4 is smooth images and less fine detail."</blockquote><hr />Note: All Prior versions have been archived.

Krea2 FinalCut

krea2_finalcut_turbo_V3_nvfp4.safetensors

krea2_finalcut_turbo_V3_int8_convrot.safetensors

Comfy_Quant_V3

Files with SHA256: 17910207ca0ead7f6839247c8285128aa353700cbfeddac9b5cfb1ebeee33c67