Ethanfel
d642255e70
Integrates GIMM-VFI alongside existing BIM/EMA/SGM models. Key feature: generates all intermediate frames in one forward pass (no recursive 2x passes needed for 4x/8x). - Vendor gimm_vfi_arch/ from kijai/ComfyUI-GIMM-VFI with device fixes - Two variants: RAFT-based (~80MB) and FlowFormer-based (~123MB) - Auto-download checkpoints from HuggingFace (Kijai/GIMM-VFI_safetensors) - Three new nodes: Load GIMM-VFI Model, GIMM-VFI Interpolate, GIMM-VFI Segment Interpolate - single_pass toggle: True=arbitrary timestep (default), False=recursive like other models - ds_factor parameter for high-res input downscaling Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
58 lines
968 B
YAML
58 lines
968 B
YAML
trainer: stage_inr
|
|
dataset:
|
|
type: vimeo_arb
|
|
path: ./data/vimeo90k/vimeo_septuplet
|
|
aug: true
|
|
|
|
arch:
|
|
type: gimmvfi_f
|
|
ema: true
|
|
modulated_layer_idxs: [1]
|
|
|
|
coord_range: [-1., 1.]
|
|
|
|
hyponet:
|
|
type: mlp
|
|
n_layer: 5 # including the output layer
|
|
hidden_dim: [128] # list, assert len(hidden_dim) in [1, n_layers-1]
|
|
use_bias: true
|
|
input_dim: 3
|
|
output_dim: 2
|
|
output_bias: 0.5
|
|
activation:
|
|
type: siren
|
|
siren_w0: 1.0
|
|
initialization:
|
|
weight_init_type: siren
|
|
bias_init_type: siren
|
|
|
|
loss:
|
|
subsample:
|
|
type: random
|
|
ratio: 0.1
|
|
|
|
optimizer:
|
|
type: adamw
|
|
init_lr: 0.00008
|
|
weight_decay: 0.00004
|
|
betas: [0.9, 0.999]
|
|
ft: true
|
|
warmup:
|
|
epoch: 1
|
|
multiplier: 1
|
|
buffer_epoch: 0
|
|
min_lr: 0.000008
|
|
mode: fix
|
|
start_from_zero: True
|
|
max_gn: null
|
|
|
|
experiment:
|
|
amp: True
|
|
batch_size: 4
|
|
total_batch_size: 32
|
|
epochs: 60
|
|
save_ckpt_freq: 10
|
|
test_freq: 10
|
|
test_imlog_freq: 10
|
|
|