Wan21_CausVid_14B_T2V_lora_rank32_v1_5_no_first_block
Back

Wan21_CausVid_14B_T2V_lora_rank32_v1_5_no_first_block
119 1 13

Other

Wan21_CausVid_14B_T2V_lora_rank32_v1_5_no_first_block

causvid 14b version 1.5.

CausVid LoRAs (Low-Rank Adapters) are experimental results extracted from the CausVid fine-tuned model, aiming to utilize the distillation effect of CausVid rather than actual causal inference capabilities.

  • v1 version: Direct extraction version, negatively affects dynamic effects, and introduces flickering artifacts at full strength.
  • v1.5 version: Similar to the above version, but removes the first module, which can eliminate flickering issues at full strength.
  • v2 version: Further pruned version, retaining only the attention layers and excluding the first module, completely resolves flickering issues and better preserves dynamic effects. It requires more inference steps and benefits from CFG (Classifier-Free Guidance).


This model is sourced from an external transfer (transfer address: https://huggingface.co/Kijai/WanVideo_comfy/tree/main ）,if the original author has objections to this transfer, you can click，
Appeal
We will, within 24 hours, edit, delete, or transfer the model to the original author according to the original author's request

user_nnr9fdwx

user_nnr9fdwx

Other

Model Information

Active
Model Type：
LoRA
Basic Model：
WAN2.1
Resource Name：
models/loras/Wan21_CausVid_14B_T2V_lora_rank32_v1_5_no_first_block.safetensors
MD5：
4fa9496aa394a5b44cc58242408b1b46

causvid 14b version 1.5.

CausVid LoRAs (Low-Rank Adapters) are experimental results extracted from the CausVid fine-tuned model, aiming to utilize the distillation effect of CausVid rather than actual causal inference capabilities.

  • v1 version: Direct extraction version, negatively affects dynamic effects, and introduces flickering artifacts at full strength.
  • v1.5 version: Similar to the above version, but removes the first module, which can eliminate flickering issues at full strength.
  • v2 version: Further pruned version, retaining only the attention layers and excluding the first module, completely resolves flickering issues and better preserves dynamic effects. It requires more inference steps and benefits from CFG (Classifier-Free Guidance).


This model is sourced from an external transfer (transfer address: https://huggingface.co/Kijai/WanVideo_comfy/tree/main ）,if the original author has objections to this transfer, you can click，
Appeal
We will, within 24 hours, edit, delete, or transfer the model to the original author according to the original author's request