Seunghyun JiandClaude Opus 4.8
2a8eb68fa5
Extracted from anima_lora/custom_nodes/comfyui-anima-pid into its own ComfyUI custom-node repo (full handoff), mirroring how the EasyControl node is distributed. Self-contained: vendors the NVIDIA PiD net under pid_net/ (Apache-2.0), bundles the fixed gemma null caption, no imports from the anima_lora library. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
22 lines
1.0 KiB
Python
22 lines
1.0 KiB
Python
"""Anima PiD ComfyUI custom nodes.
|
|
|
|
NVIDIA PiD (Pixel Diffusion Decoder) as a drop-in replacement for VAE Decode on
|
|
Anima / Qwen-Image latents: consumes a LATENT and emits a 4x super-resolved
|
|
IMAGE in one 4-step pass. The gemma text encoder is never loaded — the distilled
|
|
path has no CFG and conditions on a fixed null caption gemma(chi_prompt+"") that
|
|
ships pre-baked with the node (~1.4 MB), so there is no multi-GB text-encoder
|
|
download and no prompt input.
|
|
|
|
* ``AnimaPiDLoader`` - load a PiD qwenimage checkpoint -> ``ANIMA_PID`` socket.
|
|
* ``AnimaPiDDecode`` - ``ANIMA_PID`` + ``LATENT`` -> ``IMAGE`` (4x), with optional
|
|
latent tiling for 4K on limited VRAM.
|
|
|
|
The PiD network is vendored self-contained under ``pid_net/`` (Apache-2.0, from
|
|
nv-tlabs/PiD) — no hydra / imaginaire dependency. The PiD *weights* are NVIDIA
|
|
NSCLv1 (non-commercial) and are supplied by the user, not bundled.
|
|
"""
|
|
|
|
from .nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS
|
|
|
|
__all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS"]
|