A talking portrait needs more than plausible mouth movement. The face should remain recognizable through pauses, head turns and the end of a sentence. SoulX FlashHead is worth researching when you need an audio-driven portrait workflow, but choose a small measurable pilot before deciding how to produce a whole dialogue or presenter series.
What is SoulX FlashHead
The official repository at our inspected revision describes a 1.3B-parameter streaming portrait framework with Lite and Pro paths. Its history records February 2026 code and weight releases and March Gradio and demo updates. This guide does not claim a new October launch or evidence of a viral breakout.
The official checkpoint page separates Model_Lite and Model_Pro. Pick the checkpoint and inference path together. Do not substitute instructions for SoulX FlashTalk or LiveTalk simply because the names resemble one another.
Choose a path before installing
| Path | What to inspect first | What to record |
|---|---|---|
| Lite on one GPU | Official single-GPU Lite script and matching weights | Device, environment, input duration and saved result |
| Pro on one GPU | Official single-GPU Pro script | Actual completion time and accepted output |
| Pro on multiple GPUs | Official multi-GPU instructions and dependencies | Device count, configuration and observed behavior |
| ComfyUI | Repository-linked community node and its own instructions | Node revision, checkpoint mapping and workflow file |
The repository documents Python 3.10, a specific CUDA PyTorch build, attention dependencies and FFmpeg. Follow its current installation instructions for your machine. Published throughput figures depend on the named hardware and mode; we have not run them and do not present them as our measured speed, end-to-end latency or a minimum VRAM guarantee.
The repository links a community ComfyUI node. That is an additional integration maintained outside the core repository, so inspect its compatibility before installing. A native script guide does not prove that every node version supports both model paths or the same streaming behavior.
Prepare one portrait and a short speech pilot
Use an original fictional character or a portrait you are authorized to animate. Choose a clear head-and-shoulders composition with an unobstructed mouth. Keep the first speech test short enough to review word by word. Include a natural pause, a closed-mouth sound and a phrase ending so the test covers more than continuous open vowels.
Keep an unchanged copy of the portrait and speech file. Record the audio duration, sample format and any preprocessing. The name of a sample file is not a validation of your own recording. Listen to the actual saved audio before investigating a mouth-motion problem.
Inspect the actual Lite invocation
The pinned single-GPU Lite script calls generate_video.py with --ckpt_dir, --wav2vec_dir, --model_type lite, --cond_image, --audio_path and --audio_encode_mode stream. Inspect and replace the example input paths with your approved files in a copy of the script. Confirm output handling in the matching code before running.
Those arguments identify concrete input and model choices. They are not a generic image-to-video prompt API. We inspected the source but did not install its dependencies, download its weights or run inference. Keep that distinction when comparing this workflow with a hosted creative product.
Review the saved portrait video
Download the original portrait pilot brief and blank review ledger. Inspect the beginning, pause, phrase ending and final held frame. Check identity, mouth closure, teeth, chin, background and clothing edges. Log the timestamp and reason when an output fails.
Play the saved picture with its sound. A live preview that appears responsive does not prove the exported file has the correct start, duration or audio alignment. Also review at normal speed: a clean still frame can hide a temporal jump. Longer output requires its own sustained identity and synchronization checks.
Use the right workflow for the scene
For a simple presenter, a portrait and speech may be sufficient inputs to investigate. For a two-person drama, also plan reactions, listener coverage and room geography. A single portrait generator does not by itself assemble the conversation. The screen-direction guide and short-drama workflow help plan that edit.
For a supported hosted route, inspect Clout Image and Video and the MuseTalk workflow guide. This article does not claim that Clout runs FlashHead. Check the separate code, checkpoint and dependency terms for the intended use; a repository code license alone does not resolve every model or input permission.
