BUNNY H3 High-Dynamic Dual-Sampling Amplification Workflow.

This version primarily addresses the issues with my action LoRA—namely, noise, blurriness, and unstable details—that often occur after secondary amplification. Following specialized optimization, it now delivers clearer and more stable footage while preserving high-dynamic action.

The workflow has been optimized for 12GB of GPU memory and can achieve output quality approaching 1080p. The exact duration will vary depending on GPU memory and resolution.

Usage is simple: Enter the prompt and video duration, set the base resolution and magnification factor for the first amplification (1.5x magnification is recommended by default). If reference images are needed, load them directly; if not, bypass or disconnect the reference image node. Keep the task type set to AUTO. LoRA files and weights can be adjusted based on the subject matter; it is recommended to keep other parameters at their defaults, and do not disable the second amplification.

Additionally, please note that the hybrid model released by DaSiWa already incorporates some of my LoRA. When using his hybrid base model, there’s no need to set my LoRA weights as high as before; it’s recommended to start testing with lower weights to avoid over-emphasizing movements and visuals.

Finally, special thanks to T8 for providing the underlying fixes and dual-input logic.

I believe this is exactly how the open-source community should evolve: others use my LoRA to create better workflows, and I then perform specialized optimizations based on my understanding of LoRA’s characteristics; subsequent creators can continue to improve upon this version, driving the technology and results forward generation after generation.

I hope this workflow helps everyone generate clear, stable, high-dynamic-range H3 videos more easily.

---

BUNNY H3 High-Dynamic Dual-Sampling Amplification Workflow.

This version primarily addresses the issues with my action LoRA—namely, noise, blurriness, and unstable details—that often occur after secondary amplification. Following specialized optimization, it now delivers clearer and more stable footage while preserving high-dynamic action.

The workflow has been optimized for 12GB of GPU memory and can achieve output quality approaching 1080p. The exact duration will vary depending on GPU memory and resolution.

Usage is simple: Enter the prompt and video duration, set the base resolution and magnification factor for the first amplification (1.5x magnification is recommended by default). If reference images are needed, load them directly; if not, bypass or disconnect the reference image node. Keep the task type set to AUTO. LoRA files and weights can be adjusted based on the subject matter; it is recommended to keep other parameters at their defaults, and do not disable the second amplification.

Additionally, please note that the hybrid model released by DaSiWa already incorporates some of my LoRA weights. When using his hybrid base model, there’s no need to set my LoRA weights as high as before; it’s recommended to start testing with lower weights to avoid over-emphasizing movements and visuals.

Finally, special thanks to T8 for providing the underlying fixes and dual-input logic.

I believe this is exactly how the open-source community should evolve: others use my LoRA to create better workflows, and I then perform specialized optimizations based on my understanding of LoRA’s characteristics; subsequent creators can continue to improve upon this version, driving the technology and results forward generation after generation.

I hope this workflow helps everyone generate clear, stable, high-dynamic-range H3 videos more easily.