# 🖼️ Fix Black Image / Video Output in ComfyUI When Using SageAttention

### 1. Install ComfyUI-KJNodes

- Open **Extensions** in ComfyUI.
- Search for `ComfyUI-KJNodes`.
- Install the package.
- Restart ComfyUI after installation if required.

[![](https://wiki.chenjia.org/uploads/images/gallery/2026-05/scaled-1680-/eiQQRQX5NKJr8i51-image-1779579484588.png)](https://wiki.chenjia.org/uploads/images/gallery/2026-05/eiQQRQX5NKJr8i51-image-1779579484588.png)

### 2. Patch SageAttention Correctly

After loading your model and LoRAs, you must add the node:

`Patch Sage Attention KJ`

Then select one of the following modes:

- `sageattn_qk_int8_pv_fp16_triton`
- `sageattn_qk_int8_pv_fp16_cuda`

[![](https://wiki.chenjia.org/uploads/images/gallery/2026-05/scaled-1680-/w92rDtPATIviwJz0-image-1779579532826.png)](https://wiki.chenjia.org/uploads/images/gallery/2026-05/w92rDtPATIviwJz0-image-1779579532826.png)

### Important Notes

- The patch node should be applied **after** loading the checkpoint/model and LoRAs.
- Using the wrong SageAttention backend may result in completely black outputs.
- If one backend does not work on your system, try the other option.

### Recommended Order

```text
Load Checkpoint
    ↓
Load LoRA(s)
    ↓
Patch Sage Attention KJ
    ↓
Sampler / Generation Nodes
```

This issue is commonly related to unsupported or unstable SageAttention kernel configurations on certain GPU, CUDA, or PyTorch combinations.

### Useful Links
* [Github Issue](https://github.com/Comfy-Org/ComfyUI/issues/9184)