Before switching to Qwen (if we can fix the high-frequency artifacts and blurriness), here's my final Flux finetune extracted as a LoRA, along with a new version of my highresfix. The finetune was trained on 20k images and merged with some of my other private LoRAs, such as skin and blur fixes. It's called "Schwarzwald Klinik," which means Black Forest Clinic named after an old German TV show, and the name just fits perfectly.

This model addresses almost all my grievances with the standard dev model. It offers better prompt adherence, better text rendering, prevents style drift toward illustrations for sci-fi and fantasy prompts, fixes the "Flux chin" issue in 7 out of 10 generations, provides more diversity, and produces sharper, more detailed results than standard dev. It's also compatible with all other dev LoRas without breaking down. It's what I would have hoped Flux would be from the beginning.

The highresfix v2 is now more efficient and based on a wavelet finetune. It works effectively starting at weight 0.20 without significantly altering the generated image, aside from eliminating stripe artifacts. There are also some new nodes you can optionally use with the LoRA. You can find them in my node package via Manager or at https://github.com/42lux/ComfyUI-42lux.

Have fun! :)

Model falls under the Black Forest Dev License.

Prompts are as always from the Top Images of the last few months.