MachineDelusions/Kandinsky6-VSR-LoRas

FL Kadinsky Upscale LoRA

k6_sr_v5_x2_crisp is a rank-32 LoRA for the distilled 2-step Kandinsky 6 VSR model. It gives crisper edges and more fine texture than the released model at x2, at the cost of some faithfulness to the source.

Usage

Use it with the FL Kadinsky Upscale ComfyUI node pack. Put the file in ComfyUI/models/loras/FL_Kadinsky/:

hf download MachineDelusions/Kandinsky6-VSR-LoRas k6_sr_v5_x2_crisp.safetensors --local-dir ComfyUI/models/loras/FL_Kadinsky

Setting	Value
Scale	2x
LoRA strength	1.0 (0.7 is more natural on faces)
lq_noise_scale	0.3

Training

Trained at x2 only, on unsharp-masked 4K targets with L1, LPIPS, edge, texture and a light adversarial loss.

License

MIT, like the base model.