lora-scheduled-comfyui

LoRAScheduledTimestep

A custom ComfyUI node that applies a LoRA only during a specific window of the denoising process — and, optionally, only to one side of the CFG split — instead of throughout the entire generation. This lets you keep a base model's compositional diversity (poses, camera angles, framing) while still injecting a LoRA's style or character at the right moment. It was developed and tested with Anima / Qwen-Image DiT models, but works with any LoRA whose keys map onto the diffusion model. Applying a LoRA at full strength for the whole generation forces the model toward the LoRA's learned distribution from the very first step — which locks in composition and kills variety across seeds. But the composition of an image is decided in the early denoising steps, while style, character identity, and fine detail emerge later. By switching a LoRA on after the composition has formed, you get: - Diverse poses and camera angles (driven by the base model) - Consistent character / style (driven by the LoRA, applied late) Unlike a normal LoRA loader that merges weights, this node hooks the forward pass of the affected modules and adds the LoRA's low-rank contribution to their output, scaled by a weight that depends on the current denoising progress. The contribution is computed in fp32 for accuracy, and can be routed to only the positive or negative CFG branch (see below). If you have ComfyUI-Manager installed: 1. Open ComfyUI Manager. 2. Click Install via Git URL. 3. Paste this repository's URL: https://github.com/Murdunbad/lora-scheduled-comfyui 4. Click OK, then Restart ComfyUI when prompted…

Advanced/Lora Schedule · 1 Eingabe · 8 Parameter · 1 Ausgabe