JettHu

feat: Add transformer_function and sampler_cfg_rescaler hook

4

The authors of the [T-GATE](https://github.com/HaozheLiu-ST/T-GATE) introduced a speed up tech, like DeepCache > We find that cross-attention outputs converge to a fixed point during the initial denoising steps. > >...

Hey, found [T-GATE](https://github.com/HaozheLiu-ST/T-GATE) ([paper](https://arxiv.org/abs/2404.02747)) a technology similar to DeepCache. It brings 10%-50% speed up And is training-free and can be widely complementary to the existing studies. I have implement a...

feat: add tcd sampler

1

This PR solved #2985 . Add Trajectory Consistency Distillation Support. - Implemented a new sampling function `sample_tcd`. - TCD use `alphas_cumprod` to `calculate_denoised` & `noise_scaling`, so I add register_buffer alphas_cumprod...

Wrong sigma on mac mps backend.

1

Found a strange bug, it may be a problem with torch with mps backend. After upgrading my mac computer to the latest version. [Apple M1 Pro | 32 GB |...

Add a configured node output cache metaclass.

1

Implement a configurable node output cache metaclass to reduce unnecessary node executions. The same model currently leads to reloading due to different node IDs between workflows. Loading the model from...

Reduce repeated calls of get_immediate_node_signature in cache

3

`CacheKeySetInputSignature.get_immediate_node_signature` is called multiple times for the same node as ancestors of different nodes. Most nodes are fine, but when there are time-consuming operations in `class_def.INPUT_TYPES()`, such as `get_filename_list`, the...