Tracer / Tracer Cosplay - Flux.dev

This LoRA allows depicting a realistic Tracer / Tracer Cosplay from Overwatch in Flux.dev.

Included is only the default outfit, however it is possible (and required) to prompt for individual components of the outfit and also to have a great variance thanks to Flux.

Main trigger: tracer

Possible components:

• (open) jacket

• (gradient / yellow / orange) leggings

• pistols / pistol

• (glowing) chronal accelerator

• wearing goggles / goggles / goggles on head

• holster

Note: I sometimes misspelled chronal accelerator as chrono accelerator in the example images. This however does not seem to matter, maybe accelerator is sufficient.

Suggested LoRA weight: Depending on the style you want 0.6 – 1.0.

And with that we can talk a bit about the

Training

This is a follow up to my Mercy Outfit, I actually already collected the dataset while training the mercy LoRA, however I did not have the time so far to train on it.

As the mercy LoRA turned out fine, I used similar settings to before, with a higher learn rate and less steps, so no real new knowledge here.

The dataset is about 100 images, manually captioned in only very short captions being effectively the terms listed above.

Training Settings:

• Alpha, Dim: 16

• Total Steps: 7500

• Caption dropout: 0.05

• Resolutions: 512, 768, 1024

• Batch Size: 1

• Noise scheduler: flowmatch

• Learn rate: 4e-4 (steps 1-5000, 2e-4 steps 5000-7500)

• Linear timesteps

• Quantized (with gradient checkpointing)

Training took about 4 hours on a RTX 4090.

After training safetensor keys were converted to be compatible with Kohya, to then resize the models to rank 12 (sv_fro with 0.98).

During training my observations were that about halfway through (roughly around 4000 steps) the image quality got worse (third leg, morphed accessories, …), I however continued the training, as specific details such as the pistols were still lacking in the previews. They got better over time, however at the cost of general quality - at least in the preview images. In the actual LoRA these problems are not apparent.