toobusy DreamID-Omni Loader

Talking-head models are a whole separate dependency tree, and the toobusy pack doesn't pretend otherwise: the DreamID-Omni Loader is a wrapper that delegates to the upstream benjiyaya/ComfyUI_Dreamid-Omni nodes rather than re-implementing the engine. What it actually does for you is bundle the model-file, precision, and attention-backend choices into one node and hand you a typed pipeline handle the Talker node can consume - plus a loader_info string telling you exactly what it decided to load.

How it works

The key architectural decision is on the pack's side: it never imports or instantiates the DreamID-Omni engine itself. It looks up the upstream node classes from ComfyUI's global registry and calls them, so there's exactly one copy of the model in memory and no version skew between the two packs. If the upstream isn't installed, the node tells you so in a clear error instead of half-loading.

The other notable detail is what it deliberately leaves out. DreamID-Omni's original repo ships a cloud prompt-rewriter (dashscope) that you'd never use from ComfyUI anyway; the toobusy requirements file mirrors the upstream deps minus that package. The result is fully local - no API key, no network calls during generation.

The inputs that matter

• model_file - the DreamID-Omni checkpoint, defaulting to dreamid_omni_bf16.safetensors. This file has to be present and downloaded; the pack doesn't fetch it for you.
• precision - FP32 / BF16 / FP8. This is the VRAM lever; start at FP8 and step down if quality suffers.
• attention_backend - SDPA / Sage Attention / Flash Attention. SDPA is the safe default; switch if you have the accelerator packages installed and want the speed.

Outputs and where they go

• pipeline - the TOOBUSY_DREAMID_OMNI_PIPELINE handle, straight into the Talker node's pipeline input. There's nothing else you can do with it, and that's fine.
• loader_info - a string describing what loaded; useful as a sanity check wired into a Preview Text.

Install and gotchas

This is the "optional dependency" node, and it's a two-pack install:

python -m pip install -r custom_nodes/toobusy/requirements_dreamid_omni.txt

plus the upstream pack itself - benjiyaya/ComfyUI_Dreamid-Omni via Manager or git clone - and that pack's model files. toobusy delegates, so every model and dependency the upstream needs is a dependency of yours. The requirements file is heavy on purpose (diffusers, optimum[quanto], open-clip, moviepy, pydub, librosa), so expect a chunky install and a first-load that's slow.

Where people get burned: they install toobusy, see the DreamID nodes, and assume the talking-head model comes along. It doesn't - the upstream pack and its weights are on you. And if the loader throws a "delegated node not available" error, that's the upstream missing, not this node broken. Check Manager's installed list, restart, and the Talker side of this pair will light up.