ComfyUI-Gradual-IC-LoRA Enables Evolving LoRA Effects in AI Video

ComfyUI-Gradual-IC-LoRA is a new ComfyUI node pack that lets you gradually change the strength of an IC-LoRA effect over the course of a generated video. Instead of keeping a LoRA’s influence locked at one level, this tool fades the effect in or out at different frames without needing to touch model weights. It works by inserting tiny adapters into all 480 projection layers the LoRA touches, multiplying each layer’s LoRA contribution by a per‑frame mask that never gets baked into the model.

Burgstall‑labs built the pack after wondering what would happen if an IC‑LoRA could evolve across a clip rather than sit static. The project is an experiment with LTX 2.3 models and took days of code wrangling to get working. A major challenge was overcoming the model’s strong bias toward temporal consistency, which tries to keep every frame looking like the same scene.

How the gradual mask system works

Video creators experimenting with IC‑LoRA models can use this to direct how an effect shifts over time, like a subject slowly changing size or expression. The tool helps avoid sudden visual jumps and opens up more storytelling possibilities in short AI clips.

Developer notes and known quirks

The creator points out that the model tends to dilute extreme changes, effectively pulling a clip toward a middle ground because it wants to preserve a coherent scene. Better results come when the subject is already moving, since motion gives the model room to accept gradual adjustments. Future testing with community‑supplied IC‑LoRAs is welcomed, and the developer is eager to see how others work around the temporal‑consistency bias.

“The node intercepts every projection layer the LoRA touches (all 480 of them) and wraps each one with a tiny adapter.” — Source: Reddit