Wan Animate Podcast 2-Step LoRA
Private checkpoint upload for the accepted Wan Animate podcast-scene 2-step adapter.
Artifact
- File:
wan_animate_podcast_2step_step27_lightx2v.safetensors - SHA-256:
5f3486043a04374ef13b669b6bf9cb5a8f9897cfd91e62f2d61ca51c0961b1d6 - Base model:
Wan-AI/Wan2.2-Animate-14B - Runtime: LightX2V Wan Animate
- Schedule:
1000 -> 500 -> 0 - Quantized DiT runtime used locally: INT8-vLLM blocks with SageAttention2
- CFG: disabled
Local Acceptance Notes
The uploaded adapter is the step-27 two-step checkpoint selected against the fixed 5-second holdout.
Checkpoint-selection metrics versus the public 4-step teacher:
| Checkpoint | SSIM | PSNR |
|---|---|---|
| continuation step 9 | 0.8294 | 26.89 dB |
| continuation step 18 | 0.8417 | 27.22 dB |
| continuation step 27 | 0.8758 | 29.05 dB |
The 251-frame held-out validation generated in 158.02 seconds after load, or 1.588 generated fps, on one RTX 3090.
Included Files
wan_animate_podcast_2step_step27_lightx2v.safetensors: accepted deployable LightX2V LoRA adapterconfigs/: local runtime configs that reference this adaptervalidation/: raw local benchmark JSONs from checkpoint selection and final 16.7-second validation
This repository contains the adapter only, not the base Wan Animate model or local INT8 DiT block artifacts.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support