ComfyUI InfiniTalk AutoScale

A custom ComfyUI node that automatically scales Wan2.1 InfiniTalk video generation to any audio duration — no manual segment chaining required.

The Problem

InfiniTalk generates video in fixed-length segments (~3.2s each at 81 frames). For longer audio, you must manually chain multiple WanInfiniteTalkToVideo nodes with previous_frames connections, duplicate all sampler nodes per segment, and wire everything correctly. A 30-second video needs ~10 segments = ~60 nodes.

The Solution

One node. Any duration. Drop in your audio and image — the node reads the audio length, calculates how many segments are needed, and loops internally with proper motion frame overlap.

[LoadImage] ─────────┐
[LoadAudio] → [AudioEncoderEncode] ─┐
[UNETLoader] → [LoRA] ─┐            │
[ModelPatchLoader] ──────┤            │
[CLIPTextEncode] ────────┤            │
[ConditioningZeroOut] ───┤            │
[VAELoader] ─────────────┤            │
                         ▼            ▼
                 [InfiniTalk AutoScale]
                         │ IMAGE
                         ▼
                 [CreateVideo] → [SaveVideo]

Installation

Option 1: Clone into custom_nodes

cd ComfyUI/custom_nodes
git clone https://github.com/Biyikgokhan/ComfyUI-InfiniTalk-AutoScale.git

Option 2: Manual

Download __init__.py and node.py into:

ComfyUI/custom_nodes/ComfyUI-InfiniTalk-AutoScale/

Then restart ComfyUI. The node appears under video/infinitalk in the node search.

Required Models

These are standard Wan2.1 InfiniTalk models — no extra downloads:

Model	Purpose
Wan2_1-I2V-14B-480p (fp8 or bf16)	Base diffusion model
wan2.1_infiniteTalk_multi_fp16.safetensors	InfiniTalk adapter (model patch)
wav2vec2-chinese-base_fp16.safetensors	Audio encoder
Wan2_1_VAE_bf16.safetensors	VAE
umt5_xxl_fp8_e4m3fn_scaled.safetensors	CLIP text encoder
lightx2v_I2V_14B_480p_cfg_step_distill (optional)	Speed LoRA (4 steps instead of 30+)

Node Parameters

Parameter	Default	Description
frame_window	81	Frames per segment. Must be 4n+1. 81 = ~3.2s per segment
motion_frame_count	9	Overlap frames between segments for smooth transitions
max_segments	0	0 = auto (match audio length), or set a hard limit
steps	4	Sampling steps (4 with LightX2V LoRA, 20-30 without)
cfg	1.0	CFG scale (1.0 for distilled, higher for full model)
sampler_name	euler	Sampler algorithm
scheduler	normal	Noise schedule
seed	0	Random seed (each segment uses seed + segment_index)
audio_scale	1.0	Audio influence strength on lip sync

How It Works

1. Reads audio encoder embeddings and calculates total frame count
2. Determines segment count: ceil((total_frames - frame_window) / (frame_window - motion_frames)) + 1
3. For each segment:
   • Encodes start image (segment 1) or motion frames from previous segment
   • Windows the audio embeddings for the current time range
   • Patches the model with InfiniTalk cross-attention and outer sample wrappers
   • Runs the full sampling loop (noise → denoise → latent)
   • VAE decodes to pixel frames
   • Trims overlap and accumulates
4. Concatenates all segments into a single IMAGE output

Duration Table

Audio Length	Segments	Video Duration
~3s	1	3.2s
~6s	2	6.1s
~10s	3	9.0s
~15s	5	14.7s
~30s	10	29.2s
~60s	20	58.1s

Example Workflow

Import example_workflow.json into ComfyUI (drag & drop onto canvas). Replace your_face_image.png and your_voiceover.mp3 with your own files.

Compatibility

• ComfyUI 0.18+ (tested on 0.18.2)
• Wan2.1 InfiniTalk built-in nodes (ships with ComfyUI)
• Works with or without LightX2V distill LoRA
• Tested on RTX 5090 (32GB VRAM), should work on 24GB+ cards

License

MIT

Credits

Built on top of the Wan2.1 InfiniTalk implementation in ComfyUI core (comfy_extras.nodes_wan).