LTX-2.5 INT8 AllInOne — Speed / Quality - v1.0 — INT8 Stage 2 Speed Showcase

## LTX-2.5 INT8 All-in-One ComfyUI Workflow

An optimized LTX-2.5 INT8 workflow supporting:

- Text to Video

- Image to Video

- First & Last Frame to Video

- User-controlled resolution, FPS, duration, and exact frame count

- Optional standard LTX transformer LoRA

- Two selectable Stage 2 attention modes

### Stage 2 Modes

Speed — Sage + Sol

Recommended for 12–16 GB VRAM targets. Uses approximate attention during both stages to significantly reduce the cost of long, high-resolution generations.

Quality — Native Exact

Uses native dense attention during Stage 2 for maximum detail retention. Recommended for 24 GB+ VRAM. Long 1080p clips may still require offloading and can be substantially slower.

### Tested System

- NVIDIA RTX 5080 16 GB

- Windows

- ComfyUI

- LTX-2.5 distilled INT8 ConvRot transformer

- 1920×1088 output

- Tested up to 15 seconds

### Notes

- The 12 GB target remains experimental and has not been validated on a physical 12 GB GPU.

- Speed mode trades some fine-detail accuracy for substantially faster Stage 2 processing.

- Quality mode retains exact Stage 2 attention but its cost increases rapidly with resolution and duration.

- Model weights are not included.

- Install the attached Additional Components ZIP into the ComfyUI root, then restart ComfyUI.

- Required model names and dependencies are documented inside the workflow and component package.

Created as an independent community optimization for LTX-2.5.