woctordho

See https://github.com/woct0rdho/transformers-qwen3-moe-fused

See https://github.com/woct0rdho/transformers-qwen3-moe-fused

While the latest models are getting larger, let's not forget the old technique of ControlLoRA (LoRA version of ControlNet). I've converted some SDXL ControlNets to ControlLoRAs, which help save a few GB VRAM. We can extract a ControlLoRA from a ControlNet and a base model using `ControlLoraSave` node in my forked stability-ComfyUI-nodes. The ControlLoRA can be loaded in `Load ControlNet Model` node and used in `Apply ControlNet` node, like the original ControlNet, see the example workflow. The ones with `sdxl` in the filenames use SDXL 1.0 Base as the base model, `noob` use NoobAI-XL epsilon-pred 1.1, `illustrious` use Illustrious XL 1.0 . You may choose a base model similar to your model. The ones with `r256` use LoRA rank 256. Some simple ControlNets like Canny may still produce good results at lower ranks such as 16. The ones with `fro0.8` are further pruned with 80% Frobenius norm retained, using `resizelora.py` in Kohya's sd-scripts. They're smaller without losing much quality. Besides saving memory, this kind of pruning also works as regularization. To get results closer to the original ControlNet, you may increase the strength in `Apply ControlNet` node, such as 1.5 or 2. Update: The original stability-ComfyUI-nodes clamps the LoRA weights to 99% quantile, which causes some quality loss, see https://github.com/Stability-AI/stability-ComfyUI-nodes/issues/11 . I've made a fork to fix this. For Flux ControlLoRAs, see https://huggingface.co/woctordho/flux-control-lora

Some LoRAs pruned using `resizelora.py` in Kohya's sd-scripts. Their sizes are greatly reduced to help save VRAM. Pruning also roughly shows how much information the LoRA has learned. For two LoRAs with the same rank and fro, the larger one has more information. See PR 2194 for a fix to further reduce the LoRA size, and PR 2240 to hugely reduce the time to prune a LoRA.