# LTX 2.5 INT8 Generative Video Upscaler — Clean ComfyUI Workflow
A compact, independently assembled ComfyUI workflow for generatively upscaling an existing video with Lightricks' official LTX 2.5 Pixel Spatial Upscaler IC-LoRA and the INT8 ConvRot transformer.
## Highlights
- Video input only — no separate image input
- No prompt or text encoder at runtime
- Preserves the source audio
- Automatic source size, FPS, frame-count, and audio analysis
- 1.5x / 2.0x / 2.5x / 3.0x scale dropdown
- Automatic output sizing to multiples of 64
- INT8 ConvRot transformer
- Official Pixel Spatial Upscaler x2 IC-LoRA
- 8 distilled steps, CFG 1, Euler ancestral
- SageAttention and tiled decode for practical VRAM use
- Clean grouped UI with the implementation hidden in subgraphs
2.0x is the adapter's trained target. The other scale options are experimental canvas sizes using the same x2 adapter; larger values may invent detail and need more VRAM/time.
## Required downloads
Models are not included. Install the official LTX 2.5 INT8 transformer, Conv video VAE, audio VAE, and Pixel Spatial Upscaler IC-LoRA from:
- https://huggingface.co/Lightricks/LTX-2.5
- https://huggingface.co/Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler
Custom-node dependencies:
- https://github.com/Lightricks/ComfyUI-LTXVideo
- https://github.com/kijai/ComfyUI-KJNodes
Copy the supplied ltx25_int8_upscale_core folder into ComfyUI/custom_nodes, restart, then open the JSON.
## Notes
- A source frame count of 8n+1 is recommended.
- Use a source with an audio track if you want audio preserved.
- No LTX text encoder is required at runtime; the package includes a fresh empty-prompt conditioning cache generated from the official LTX 2.5 INT8 encoder.
- The release was checked for graph loading and node availability on RTX 5080 16 GB. The newly assembled graph was not rendered during packaging.
## License
LTX 2.5 and the Pixel Spatial Upscaler are by Lightricks. This download includes the complete LTX-2.x Community License and a notice describing the modifications. Review the license before use or redistribution.
Description
Initial public release.
- Clean video-only ComfyUI workflow
- INT8 ConvRot transformer
- Official LTX 2.5 Pixel Spatial Upscaler x2 IC-LoRA
- 1.5x, 2.0x, 2.5x, and 3.0x scale dropdown
- Preserves source audio
- No prompt or text encoder required at runtime
- 8 distilled sampling steps, CFG 1
- SageAttention and tiled VAE decoding
- Model weights are not included
The 2.0x option is the adapter's trained target. Other scale values are experimental.