Models difference

#3
by Problemass - opened

There are 2 models: "flux1-krea-dev_float8_e4m3fn_learned_svd.safetensors" and "flux1-krea-dev_float8_e4m3fn_scaled_learned_stochastic.safetensors". What's the difference between them and which one will be better in terms of realism?

The SVD version should perform better in most if not all scenarios. The stochastic version uses stochastic rounding, whereas the svd version uses TPEC-Quant to determine the best rounding.

Sign up or log in to comment