mignonjia/LongLive-1.0-Stage-1-Diffusers

LongLive 1.0 Stage 1 Diffusers

Diffusers-style LongLive stage-1 checkpoint converted from:

/mnt/weka/home/hao.zhang/mhuo/FastVideo/LongLive/longlive_models/models/longlive_base.pt

The model uses the causal Wan 2.1 T2V 1.3B layout. The LongLive weights replace the transformer checkpoint; scheduler, tokenizer, text encoder, and VAE follow wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers.

Recommended runtime kwargs:

model_kwargs:
  timestep_shift: 5.0
  local_attn_size: 12
  sink_size: 3

FastVideo training config equivalent:

models:
  student:
    _target_: fastvideo.train.models.wan.WanCausalModel
    init_from: mignonjia/LongLive-1.0-Stage-1-Diffusers
    trainable: true

pipeline:
  flow_shift: 5
  dit_config:
    local_attn_size: 12
    sink_size: 3