| --- |
| license: mit |
| --- |
| tags: |
| - text-to-video |
| - video-to-video |
| - comfyui |
| - ltx-video |
| - ltx-2.3 |
| - character-consistency |
| - temporal-coherence |
| --- |
|
|
| # LTX-2.3 Face Preservation & Likeness Continuity Pipeline |
|
|
| An advanced ComfyUI workflow architecture designed to achieve flawless character identity preservation across extreme temporal motion sequences (including full 360° rotations and out-of-sight blind spots) using a unified **Custom LoRA + IC-LoRA Spatial-Temporal Anchor** system. |
|
|
| --- |
|
|
| ## 🚀 The Core Challenge |
|
|
| Maintaining a subject's likeness in diffusion-based video models during major structural changes—like a person turning fully around—typically leads to identity drift, facial morphing, or "hallucinated" features when the face returns to the camera view. |
|
|
| By decoupling conditioning, this pipeline splits identity retention into a micro and macro layer: |
| 1. **Micro Structural Fidelity:** A custom identity LoRA handles fine-grain muscle memory and high-fidelity texture replication. |
| 2. **Macro Spatial Anchoring:** An LTX Image-Conditioning LoRA (IC-LoRA) Guide serves as a continuous cross-attention homing beacon across the global frame context. |
|
|
| --- |
|
|
| ## 🏗️ Architecture & Pipeline Topography |
|
|
| ```text |
| STAGE 1: Native Spatial Generation (512x512) |
| [Base Model + Custom LoRA] |
| │ |
| ▼ |
| [LTXAddVideoICLoRAGuide] ───► [KSampler] ───► [Audio Latent Injection] ───► [VAE Decode] |
| (Pure Video Latents Only) |
