Anima Turbo gets an INT8 build for faster local anime images
A new Nunchaku checkpoint makes Anima Turbo v1.1 available with INT8 weights and activations. Its creator reports lower memory use and faster runs on three NVIDIA GPUs.
A new route to local generation
A new community build of Anima Turbo v1.1 puts its image-generation transformer into Nunchaku’s INT8 weight-and-activation format. The Hugging Face repository, created October 13, now contains the checkpoint, supporting components, setup instructions and paired BF16-versus-INT8 sample images. Anima comes from CircleStone Labs and Comfy Org and targets anime and illustration rather than photorealism. This is a quantization of its distilled Turbo variant, not a new base model from those developers.
What the creator measured
The author reports a 1.36-times speedup on an RTX 3090, 1.25 times on a 4090 and 1.23 times on a 5090 for 1024-pixel images at eight steps. In the same project tests, peak allocated GPU memory fell from 9.44 GiB to 7.82 GiB. AITrending has not rerun those benchmarks or judged image quality from the provided samples. The card says text encoders and the VAE remain in BF16, so the memory saving is narrower than quantizing the entire pipeline. Its current installation instructions call for a specific Linux, Python 3.11 and CUDA 13 wheel plus a compatible Diffusers build. The weights also inherit Anima’s non-commercial license; the upstream card distinguishes model-use restrictions from rights in generated images.