A digital human demo is easy: one portrait, one audio clip, one talking head. A digital human tool has to survive a script — multiple shots, multiple images, loops that continue rather than restart. This release is about that gap.
Two graphs, both large:
| Graph | Nodes |
|---|---|
| Single-image loop digital human | 210 |
| Multi-image shot-switching digital human | 197 |
Both load WanVideoVAELoader → Wan2_1_VAE_bf16.safetensors in bf16. That is the Wan 2.1 VAE, which tells you the conditioning lineage here is Wan-family video rather than a pure LTX path.
Why 200 nodes for a talking head
The node count is the honest answer to "what breaks in production". A single-shot talking head needs maybe 40 nodes. The extra 160 in these graphs are doing the things a demo skips:
- Shot continuity — the subject has to survive a cut. Multi-image switching means the graph carries identity across a change of reference, not just a change of frame.
- Loop continuation — a loop that ends and restarts shows a seam. Continuing means the tail of one segment conditions the head of the next.
- Audio length adaptation — the audio decides the frame count, and the graph has to reconcile a fixed latent geometry against variable-length speech.
- Resolution handling — 720p output from a base that generates smaller requires a deliberate path, not a resize.
The two-graph split is the real design decision
Single-image loop (210 nodes) is for a long take where the subject does not change. Everything in the graph is optimised for continuity across the segment boundary.
Multi-image shot-switching (197 nodes) trades loop continuity for reference flexibility. It has fewer nodes, which is initially surprising — switching references means less carry-forward machinery, because each shot gets a fresh anchor rather than a continuous one.
Choosing between them is a content question, not a quality question:
- Monologue, long take, one person → single-image loop
- Dialogue, multiple angles, reaction shots → multi-image switching
Running the multi-image graph for a monologue wastes its flexibility and loses the loop continuity you actually wanted.
Where this class of pipeline still hurts
- Identity across a cut is the weakest point. A hard cut with a new reference is where drift is most visible, because the viewer has no continuous motion to reassure them.
- Lip sync under 720p — at 720p the mouth region is a small number of pixels, so sync errors read as softness rather than misalignment.
- Long audio — the loop machinery helps, but a five-minute script is many segments, and each boundary is a place where a small error compounds.
Practical starting point
Do not start with the 200-node graph. Render one clean 10-second talking head first, confirm the identity holds, and only then move to the loop or switch graph. Debugging a bad anchor inside 200 nodes is much slower than debugging it inside 40.
Workflows
Two graphs, published on RunningHub with invite code rh-v1111:
- LongCat Avatar multi-image shot-switching digital human
- LongCat Avatar single-image loop digital human
Files
The ComfyUI workflow exports are in the companion repository:
https://huggingface.co/AI-KSK/ksk-video-251-comfyui-workflows
Video walkthrough (Chinese): https://youtu.be/jTgtq_YCBNU
Bilibili: https://www.bilibili.com/video/BV1gdG161EYB/
Notes
- Workflow files only. Model weights are not included; each model keeps its own licence.
- These are tutorial exports, not a packaged product.
- Node counts quoted above are read from the published JSON files.
- Independent AI-KSK release. Not affiliated with the model authors or RunningHub.
Optional API platform - 88API (affiliate link)
Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.
https://88api.ai/sign-up?aff=usx3
Disclosure: I may earn a commission from eligible purchases after you register through this referral link.
Top comments (0)