blacksnowskill/anima_booster

ANIMA_BOOSTER, also called BSS, is an add-on package for ComfyUI, a popular tool for running image generation models on your own computer. It targets one specific model called Anima DiT 2B and tries to make it run faster while using less video memory on NVIDIA graphics cards. The author says the combined speed-up is between 3.5 and 5.0 times the speed of the default Anima workflow, with no visible drop in image quality. The package bundles three main pieces. The first is a one-click toggle for torch.compile, a PyTorch feature that compiles parts of the model the first time you run it so later runs go faster, claimed at 20 to 40 percent. The second is SageAttention, a faster 8-bit version of a calculation that image models do a lot of. The third is something the author calls Adaptive TeaCache, which skips repeat calculations in later steps of the image generation when the picture is mostly finished and small changes do not matter as much. Installation is described two ways: through the built-in ComfyUI Manager by pasting the GitHub URL, or by cloning the repository into the custom_nodes folder by hand. The README also has a long section about Windows users who run the portable ComfyUI build, because they need precompiled .whl files for SageAttention and Triton instead of the normal pip install, which would fail without a C++ compiler installed. Once installed, three new nodes appear under the BSS/AnimaBooster category: a model loader, a TeaCache node with two version modes, and a latent image utility that keeps image sizes aligned to the model's grid. The README is clear that the optional libraries are optional, and the nodes will quietly turn off features that need missing libraries rather than crashing.