Ethanfel/Comfyui-STAR
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Revert model loading to original HF-based paths

Browse Source 
Reverts text encoder and VAE loading back to using HuggingFace preset
names / repo IDs (downloading to library cache) while keeping the
attention dispatcher improvements (4D SDPA, math backend).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
Ethanfel
2026-02-15 01:20:37 +01:00
parent 4c6c38f05a
commit f03c4853f1
2 changed files with 7 additions and 96 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
inference.py
View File

@@ -294,23 +294,7 @@ def load_model(model_path: str, precision: str, offload: str, device: torch.devi
from video_to_video.utils.config import cfg
print("[STAR] Loading text encoder (OpenCLIP ViT-H-14)...")
_models_dir = (SCRIPT_DIR / ".." / ".." / "models").resolve()
_te_filename = "open_clip_vit_h_14_laion2b.bin"
_te_dir = _models_dir / "text_encoders"
_te_path = _te_dir / _te_filename
if not _te_path.is_file():
from huggingface_hub import hf_hub_download
_te_dir.mkdir(parents=True, exist_ok=True)
print(f"[STAR] Downloading OpenCLIP ViT-H-14 text encoder to {_te_dir}...")
_dl = hf_hub_download(
repo_id="laion/CLIP-ViT-H-14-laion2B-s32B-b79K",
filename="open_clip_pytorch_model.bin",
local_dir=str(_te_dir),
)
shutil.move(str(_dl), str(_te_path))
print(f"[STAR] Text encoder saved to {_te_path}")
text_encoder = FrozenOpenCLIPEmbedder(device=device, pretrained=str(_te_path))
text_encoder = FrozenOpenCLIPEmbedder(device=device, pretrained="laion2b_s32b_b79k")
text_encoder.model.to(device)
negative_y = text_encoder(cfg.negative_prompt).detach()
text_encoder.model.to(keep_on)
@@ -339,33 +323,9 @@ def load_model(model_path: str, precision: str, offload: str, device: torch.devi
from diffusers import AutoencoderKLTemporalDecoder
print("[STAR] Loading temporal VAE...")
_vae_dir_name = "svd-temporal-vae"
_vae_path = _models_dir / "vae" / _vae_dir_name
if not (_vae_path / "config.json").is_file():
from huggingface_hub import hf_hub_download
_vae_path.mkdir(parents=True, exist_ok=True)
print(f"[STAR] Downloading SVD temporal VAE to {_vae_path}...")
for _f in ["config.json", "diffusion_pytorch_model.fp16.safetensors"]:
_dl = hf_hub_download(
repo_id="stabilityai/stable-video-diffusion-img2vid",
subfolder="vae",
filename=_f,
local_dir=str(_vae_path),
)
_dest = _vae_path / _f
if Path(_dl) != _dest and Path(_dl).is_file():
shutil.move(str(_dl), str(_dest))
# Clean up empty vae/ subdirectory created by hf_hub_download
_vae_sub = _vae_path / "vae"
if _vae_sub.is_dir():
try:
_vae_sub.rmdir()
except OSError:
pass
print(f"[STAR] Temporal VAE saved to {_vae_path}")
vae = AutoencoderKLTemporalDecoder.from_pretrained(
str(_vae_path), variant="fp16",
"stabilityai/stable-video-diffusion-img2vid",
subfolder="vae", variant="fp16",
)
vae.eval()
vae.requires_grad_(False)

57
nodes.py
View File

@@ -223,25 +223,8 @@ class STARModelLoader:
# ---- Text encoder (OpenCLIP ViT-H-14) ----
from video_to_video.modules.embedder import FrozenOpenCLIPEmbedder
_te_filename = "open_clip_vit_h_14_laion2b.bin"
_te_path = folder_paths.get_full_path("text_encoders", _te_filename)
if _te_path is None:
from huggingface_hub import hf_hub_download
import shutil
_te_dir = folder_paths.get_folder_paths("text_encoders")[0]
os.makedirs(_te_dir, exist_ok=True)
print(f"[STAR] Downloading OpenCLIP ViT-H-14 text encoder to {_te_dir}...")
_dl = hf_hub_download(
repo_id="laion/CLIP-ViT-H-14-laion2B-s32B-b79K",
filename="open_clip_pytorch_model.bin",
local_dir=_te_dir,
)
_te_path = os.path.join(_te_dir, _te_filename)
shutil.move(_dl, _te_path)
print(f"[STAR] Text encoder saved to {_te_path}")
text_encoder = FrozenOpenCLIPEmbedder(
device=device, pretrained=_te_path
device=device, pretrained="laion2b_s32b_b79k"
)
text_encoder.model.to(device)
@@ -281,42 +264,10 @@ class STARModelLoader:
# ---- Temporal VAE (from HuggingFace diffusers) ----
from diffusers import AutoencoderKLTemporalDecoder
_vae_dir_name = "svd-temporal-vae"
_vae_path = None
for _d in folder_paths.get_folder_paths("vae"):
_candidate = os.path.join(_d, _vae_dir_name)
if os.path.isfile(os.path.join(_candidate, "config.json")):
_vae_path = _candidate
break
if _vae_path is None:
from huggingface_hub import hf_hub_download
import shutil
_vae_base = folder_paths.get_folder_paths("vae")[0]
_vae_path = os.path.join(_vae_base, _vae_dir_name)
os.makedirs(_vae_path, exist_ok=True)
print(f"[STAR] Downloading SVD temporal VAE to {_vae_path}...")
for _f in ["config.json", "diffusion_pytorch_model.fp16.safetensors"]:
_dl = hf_hub_download(
repo_id="stabilityai/stable-video-diffusion-img2vid",
subfolder="vae",
filename=_f,
local_dir=_vae_path,
)
_dest = os.path.join(_vae_path, _f)
if _dl != _dest and os.path.isfile(_dl):
shutil.move(_dl, _dest)
# Clean up empty vae/ subdirectory created by hf_hub_download
_vae_sub = os.path.join(_vae_path, "vae")
if os.path.isdir(_vae_sub):
try:
os.rmdir(_vae_sub)
except OSError:
pass
print(f"[STAR] Temporal VAE saved to {_vae_path}")
vae = AutoencoderKLTemporalDecoder.from_pretrained(
_vae_path, variant="fp16",
"stabilityai/stable-video-diffusion-img2vid",
subfolder="vae",
variant="fp16",
)
vae.eval()
vae.requires_grad_(False)