The engines.
Throughput and peak memory measured on the serving GPU, not estimated. Every engine here is licensed for commercial use; the studio's modes run on them.
Centshot/sdxl-base-fp8
TensorRT FP8 on Ada/Hopper. Fits comfortably, fastest base.
16.4 steps/s9.6 GBq92~1.7s
text-to-image
Centshot/sdxl-lora-fp8
The LoRA style fused into an FP8 engine. The cheapest way to ship the brand look.
15.8 steps/s9.9 GBq90~1.8s
text-to-image
Centshot/sdxl-base-int8
TensorRT INT8, entropy-calibrated. Near-FP16 quality, ~45% lighter.
12.6 steps/s11.2 GBq95~2.4s
text-to-image
Centshot/schnell
FLUX.1-schnell quantised to FP8 on Ada tensor cores. Guidance is ignored: the model is guidance-distilled.
not benchmarked yet
text-to-image
Centshot/qwen-image
20.4B rectified-flow transformer with true classifier-free guidance and text legible enough to read back. The first model here whose quality metric can actually fail.
not benchmarked yet
text-to-imagecoming soon