Major Changes: - Migrate all nodes to ComfyUI V3 schema (stateless design, new IO types) - Fix GGUF weight caching VRAM leak (non-persistent buffers + _apply override) - Fix GGUF torch.compile compatibility (@torch._dynamo.disable on dequant) - Centralize compatibility checks (Flash/Triton/GGUF/Conv3d in compatibility.py) - Make flash_attn optional with graceful SDPA fallback - Add attention_mode UI option (sdpa/flash_attn) to DiT loader - Fix color correction batch padding error (trim input_video consistently) Code Quality: - Remove internal_execute for clarity (stateless node design) - Add startup logging for optimization status - Improve error messages with installation instructions - Add get_unique_id() for V3 node-specific caching - Standardize parameter names (dit_cache/vae_cache)
8 lines
194 B
Python
8 lines
194 B
Python
"""
|
|
ComfyUI-SeedVR2_VideoUpscaler
|
|
Official SeedVR2 integration for ComfyUI
|
|
"""
|
|
|
|
from .src.interfaces import comfy_entrypoint, SeedVR2Extension
|
|
|
|
__all__ = ["comfy_entrypoint", "SeedVR2Extension"] |