Questions on Reproducing Phased Consistency Models (PCM)

[ouroboros-phy]

Thank you for your excellent work on Phased Consistency Models (PCM). While attempting to reproduce the results from your paper, I encountered some issues:

1. In PCM, the ODE trajectory is partitioned into m sub-trajectories based on the multiphase hyperparameter. When multiphase=1, the model effectively becomes a single-step generator. I am confused whether the pretrained lcmlike model correspond to a single-phase PCM implementation?

2. During reproduction, I observed that the quality of generated results is highly sensitive to the CFG value.Would it be appropriate to use CFG=1 for the smallcfg model and CFG=7.5 for the normalcfg model, as these values appear to be common defaults? Are there any recommended adjustments for optimal performance?

3. The paper adopts FID-CLIP for evaluation, building upon previous works. However, I noticed that implementation details are sparse even in these references. I wonder if the metric is computed by simply replacing the Inception network with a CLIP feature extractor and which CLIP variant was employed (e.g., ViT-B/32, ViT-L/14).