iAskAsk Local FlashHead Loader

This node is the entrance fee. The "iAskAsk Local FlashHead Loader" is where the whole talking-head workflow's cost shows up: it drags SoulX-FlashHead 1.3B - about 12GB of weights - off your disk and onto your GPU. It does no animation itself. It's the one-time setup that hands the downstream sampler a ready-to-run pipeline, and it's also where most people bounce off this pack, because it's the node that decides whether your hardware is up to the job at all.

FlashHead, for the uninitiated, is Soul-AILab's audio-driven talking-head model: you feed it one portrait and a voice clip, and it moves the face to match the speech. It's the newer, smaller entrant in a niche the KB tracks as "you need a face to move to a voice" - a corner where InfiniteTalk and LTX-2.3's native audio are the usual names, and where everything runs locally but nothing is as simple as loading a checkpoint. This pack wraps it as two typed nodes, and this is the first one. The one real piece of community signal for SoulX-FlashHead is someone planning to run it on a 12GB 4070 Super for a music video - which is about the honest spec floor.

What it actually loads

Under the hood the node builds a FlashHeadPipeline pointed at a checkpoint directory, and here's the part worth knowing before you download anything: the pack ships two model variants in the same folder. model_type picks between them, and they use different video VAEs:

• lite (default) - reads the Model_Lite + VAE_LTX subdirectories and pairs the transformer with an LTX VAE. Lighter, faster, and the author's default.
• pro - reads Model_Pro + VAE_Wan/Wan2.1_VAE.pth and uses the Wan 2.1 VAE. The heavier, quality-side path.

Both variants share the same Wan-based audio-driven transformer, so you're not choosing between architectures so much as between a smaller/faster pairing and a bigger one. Weights load in bfloat16 on Ampere-and-newer GPUs, falling back to float16 otherwise, and the loader hard-codes a single-GPU setup - no multi-GPU sharding to worry about.

The output is a typed FlashHead Pipeline object (labeled "FlashHead Pipeline"). It has exactly one consumer: the iAskAsk Local FlashHead Sampler in the same pack. You can't wire it into anything else, and that's by design - it's pack-internal plumbing.

The inputs that matter

There are only three, and two of them are usually left alone:

• model_type - pro or lite, default lite. Change this before you run if you want the heavier variant.
• ckpt_dir - custom path to the SoulX-FlashHead checkpoint root. Leave it empty and it defaults to ComfyUI/models/Soul-AILab/SoulX-FlashHead-1_3B/.
• wav2vec_dir - path to the wav2vec2 audio encoder. Empty defaults to ComfyUI/models/facebook/wav2vec2-base-960h/.

Installing it - and the trap in the README

Install is the usual two-step. From ComfyUI Manager, search for "ComfyUI_iAskAsk_SoulX_FlashHead" and install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/iaskask-com/ComfyUI_iAskAsk_SoulX_FlashHead
cd ComfyUI_iAskAsk_SoulX_FlashHead
pip install -r requirements.txt

The requirements pull in a genuinely heavy stack - diffusers, transformers, optimum-quanto, librosa, mediapipe, imageio-ffmpeg, av - which means real risk of version conflicts with whatever ComfyUI already has. If you hit a wall there, that's the classic custom-node dependency fight, not your fault.

Then the models, ~12GB plus a small one:

python download_model_huggingface.py      # global
python download_model_modelscope.py       # if you're in China (falls back to HF for wav2vec2)

Here's the gotcha the README glosses over: those scripts save into the pack's own models/ folder, but the loader's empty-field defaults point at ComfyUI/models/Soul-AILab/.... The two paths don't match. If you run the script and leave the fields blank, the loader will look in the wrong place and fail to find the model. Either pass the pack's models/ paths explicitly into ckpt_dir and wav2vec_dir, or move the downloaded folders into the ComfyUI models tree. This is the single most common "why won't it load" failure for this pack, and it's not documented anywhere obvious.

Other things that'll bite: forgetting the wav2vec2 encoder (it's small, but it's mandatory - the pipeline can't embed audio without it), and trying this on a card with much less than ~12GB of VRAM. First load also takes a while because nothing is cached - that slow startup is normal, not a hang.