This is a highly optimized, mixed-precision (MXFP8) quantized version of the Anima Base V1.0 model. It was created to provide a much lighter and faster generation experience without noticeably sacrificing visual fidelity.
If you are running on a GPU with limited VRAM, or if you want to free up memory to run heavier ComfyUI workflows (like multiple ControlNets, upscale pipelines, or LoRAs), this checkpoint is for you.
Performance & Stats vs Original Anima V10:
VRAM Usage: ~2.2 GB (Reduced from ~4.0 GB)
File Size: 2.15 GB (Reduced from ~4.0 GB)
Generation Speed: ~2.3 it/s (Increased from ~1.9 it/s)
Precision: Mixed-precision (
mxfp8,int8,float8)
Description
FAQ
Comments (7)
is this faster than int8 convrot?
Yes, on my RTX 5000 series it runs about 10-15% faster than INT8.
Quick comparison from my tests with a turbo lora:
• MXFP8: ~2.17 to 2.33 it/s (around 4.5s per generation)
• INT8: ~1.95 to 2.08 it/s (around 4.8s per generation)
• VRAM: Basically the same (~2.1 - 2.2 GB)
@jancok It depends on your card. My old card only has INT8 acceleration, for instance; I think in descending order of oldness it's INT8, FP8, NV4, MXFP8 for acceleration support?
@pedrodenovox549 hmm thats weird. mxfp8 is slower than int8 convrot in my rtx 5060 ti16G
@pedrodenovox549 are you using forge neo or comfyui?
@jancok ComfyUI
@jancok I tested it using the same prompt and seed in the first run to load the model, and a second run without changing anything except the seed...
I tested it on the Anima and Krea2 models.
The Anima mxfp8 models performed better than the int8 models in all cases... With LORA, without LORA, with Turbo, and without Turbo.