<h3 id="~-recommended-settings-and-training-parameters-below!-~-eai1ou6em">~ Recommended Settings &amp; Training Parameters Below! ~</h3>This is a Flux LoRA of singer/songwriter/DJ Jahan Yousaf of EDM group Krewella. I've attempted to train both her two-toned dark hair with blonde ombré (circa Get Wet) era, and the long dark hair era in one LoRA. The adherence rate for this is not 100%, but the likeness is there regardless. This was sort of a test for me to start trying to train great quality character LoRAs for Flux. Basically what I've learned from this is don't have hands in your datasets, alpha mask training is great, and upscale training data before cropping/resizing! If you have any tips on training Flux please let me know! I'm looking to improve.<h2 id="recommended-settings-l16704cvz">Recommended Settings</h2>Strength ... 1.10Flux Guidance ... 3.0 - 4.0Steps ... 40 - 45Sampler/Scheduler ... any really but Euler/Beta seems to be good.It seems to do best with shorter tag-like captioning styles focusing on describing the subject. Check the showcase for examples + ComfyUI workflow.It's been trained to use gwkr for ombré hair color and nwpo for black hair. Sometimes it may generate one or the other based on the prompt. You can always specify long black hair or dark brown hair with ombre for better adherence.<h2 id="training-parameters-78lmnsr2i">Training Parameters</h2>Because I hate gatekeeping, in the spirit of open source goodness, here's what I used ~Tool ... KohyaLR ... 0.0003Dim or "Rank" / Alpha ... 8 / 1Steps ... 2500Scheduler ... AdamWScheduleFree, Cosine22 total images, 10 repeats, Alpha Mask enabledBatch ... 2Resolution ... 1024x1024 (Upscaled with 4x-NMKD-Siax_200k or 4xFaceUpLDAT then resized)All images and the model file itself contain full inference and training metadata. 💜

[Flux] Jahan Yousaf (Krewella)

This model doesn't necessarily require any trigger words, but it helps with likeness (I think...)

Files with SHA256: 92644c97477f04f9b20f9e7b1a5ab1ffae4d5dc500cf5897c5a58ad9e082f92e