A LoRa designed to address the high-resolution checker boarding and banding issues in Flux [dev] and schnell. Initial generation up to 3MP and img2img up to 6MP. You can find the corresponding workflow here: https://civarchive.com/articles/9917

With this LoRa and my corresponding ComfyUI workflow, you can generate images directly at 3MP and apply a high-resolution fix to achieve 5MP natively, without the need for tiled scaling. On a 4090 GPU (with Torch Compile enabled), the runtime is approximately 35 seconds for 3MP and 95 seconds for 6MP.

This LoRa works best with my high-resolution inference workflow in ComfyUI, though it can likely be adapted for use in Forge as well. Flux [dev] tends to introduce artifacts when generating images at resolutions higher than 2MP. To mitigate this, we include a node in the workflow that normalizes the sampling schedule for higher resolutions.

At a low strength setting (0.15), the LoRa primarily focuses on optimizing high-resolution noise scheduling. At medium strength (0.42), it can be used for inpainting faces or other skin details to get rid of the plastic look of flux. While it can be used directly for generation, higher strength levels (0.8–1) may cause a style shift toward realism. However, it effectively eliminates the infamous “flux chin” issue.

My datasets tend to consist of more realistic images than anime or illustrations. I would love to train an anime/illustration version as well. If anyone has a good general anime/illustration dataset (1440p resolution and high-quality images) and is willing to provide it, I would train an anime/illustration version of the LoRa.