DEV Community

AI-KSK
AI-KSK

Posted on Edited on Originally published at huggingface.co

InfiniteTalk Native in ComfyUI: Looping a Talking Head to Unlimited Length

Infinite length video is not a bigger generation — it is a loop that does not show its seam. InfiniteTalk's native ComfyUI support makes that practical, and this release ships both single and dual-character loops plus a LTX-2 lip-sync path.

The four graphs

Workflow Nodes Base
FlashVSR V1.1 upscale 4 FlashVSR
LTX-2 distilled lip-sync 106 LTX-2 19B
Single-person InfiniteTalk loop 68 SkyReels V3 A2V
Dual-person InfiniteTalk loop 65 SkyReels V3 A2V

A 4-node upscale graph

FlashVSR V1.1放大 is four nodes. That is about as minimal as a working graph gets: load model, load video, upscale, save.

Worth noting because it is the counterexample to everything else in this series. Minimal graphs exist; they just do not do much. A 4-node upscaler is fine precisely because upscaling has no conditioning to manage.

The InfiniteTalk base

SkyReelsV3-A2V_fp8_scaled_mixed.safetensors — audio-to-video, FP8 scaled, mixed precision.

Supporting loaders:

  • umt5_xxl_fp8_e4m3fn_scaled (type wan) — the Wan-family multilingual encoder
  • Wan2.1_VAE.pth — the Wan VAE, in pickle format
  • lightx2v_I2V_14B_480p_cfg_step_distill_rank64_bf16 — a LightX2V distillation LoRA

The 480p in that LoRA name is deliberate: LightX2V's distilled model is trained for 480p work. Using it above its training resolution produces softness, not failure.

How the loop avoids a seam

A talking-head loop that restarts shows a jump: the mouth position at the end does not match the start. InfiniteTalk's approach carries audio state across the boundary so the model continues rather than restarts.

Two consequences follow:

The audio must be long and continuous. You cannot loop three five-second clips and expect continuity; the loop needs a single audio track to advance through.

The seam moves to wherever the audio has an edit. A pause, a breath, a cut in the audio becomes the place where the loop shows itself.

That is a meaningful difference from clip-chaining approaches, where the seam is at a fixed frame count.

The dual-person variant

65 nodes versus 68 for single — the dual version is smaller.

Same pattern as in other dual-character work: switching between references is less machinery than carrying one identity continuously. The single-person loop has more carry-forward state to maintain.

The LTX-2 lip-sync path

106 nodes, and it uses the older LTX-2 (19B) generation rather than 2.3:

  • EmptyLTXVLatentVideo at 768 × 512, 241 frames
  • LTXVEmptyLatentAudio at 121, 25, 1
  • euler_ancestral sampler (not the deterministic euler used in the InfiniteTalk graphs)

241 frames is about ten seconds — much longer than the 97-frame unit common elsewhere. And the audio latent is 121, roughly half the frame count.

Note the two DualCLIPLoader variants: DualCLIPLoaderGGUF with the fp8 gemma and DualCLIPLoader with a scaled gemma. Both present, presumably for comparison.

Where this breaks

  1. .pth VAE — tooling and trust considerations.
  2. 480p distillation LoRA caps effective resolution.
  3. Loop seam moves to audio edits — you cannot hide it with a frame-count cut.
  4. The LTX-2 lip-sync graph uses the older generation, so tuning from LTX 2.3 does not transfer.
  5. Two text encoder loaders in one graph is confusing; only one is likely active.

Practical approach

Start with the 4-node upscale to confirm FlashVSR works. Then the single-person loop on a 30-second audio track — long enough to see whether the seam appears. Only then try dual-person.

Workflows

Four graphs, published on RunningHub with invite code rh-v1111:

  • FlashVSR V1.1 upscale
  • LTX-2 distilled lip-sync
  • Single-person InfiniteTalk unlimited-length native loop
  • Dual-person InfiniteTalk unlimited-length native loop

Files

https://huggingface.co/AI-KSK/ksk-video-171-comfyui-workflows

Notes

  • Workflow files only. Model weights are not included; each keeps its own licence.
  • Tutorial exports, not a product.
  • Node counts and loader values quoted above are read from the published JSON files.
  • Independent AI-KSK release. Not affiliated with the model authors, FlashVSR 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)