Some users of the latest Fluxmania models reported unpredictable streaking and banding artifacts — particularly after merging versions that include the LoRA Fluxartis IIB (available on my page).
Interestingly, this issue seems tied to a few overdominant feature deltas within that LoRA.
A brilliant fix was suggested by a member of the Fluxmania community: applying Fluxartis IIB at a negative weight (e.g., -0.44) during inference. This acts as a corrective vector — subtracting the problematic keys and reducing artifacts — while preserving the intended stylistic imprint. It’s an elegant workaround that restores clarity without needing to retrain or remerge.
Fluxartis IIB : https://civitai.com/models/988728?modelVersionId=1539595