This isn't clickbait. After trying the INT8 Convrot model, I've decided to switch from FP8 completely.
In short: much faster, with no noticeable loss in image quality. If you're using Krea2, I highly recommend the INT8 Convrot version.
Using Krea2 INT8 Convrot only requires two steps:
1. Update ComfyUI to the latest version, as INT8 support was added recently.
2. Download the official model from Hugging Face.
Official download:
https://huggingface.co/Comfy-Org/Krea-2
Make sure you download the official model. Some unofficial versions on Hugging Face will trigger backend errors and won't work properly.
Below is a quick comparison on my laptop.
## I. Basic Info
### 1. Hardware
Same as usual: I am still using my old low-spec laptop. The basic setup is:
- CPU: Intel Core Ultra 5 225H
- RAM: 32GB
- GPU: NVIDIA GeForce RTX 5060 Laptop
- VRAM: 8GB
### 2. Models
- krea2turbofp8_scaled.safetensors (FP8)
- krea2turboint8_convrot.safetensors (INT8)
### 3. Generation Settings
- 1024 × 1024
- 8 steps
- euler, simple
II. Speed Comparison
| Model | Generation Speed | Total Time |
|---|---|---|
| FP8 | 2.91 s/it | 23 s |
| INT8 | 1.60 s/it | 12 s |
The results above are averaged over multiple single-image generations. INT8 cuts the total generation time by nearly 50% compared to FP8.
III. No Quality Loss
Note: The first set uses my newly released LoRA, # VG28 Eira , The second set uses no LoRA.
As you can see, FP8 and INT8 produce slightly different details, but neither is objectively better. Image quality is essentially the same.
Finally, my previously released # Basic Workflow: Krea 2 T2I v1.0 works with the INT8 model out of the box, so no changes are required.
Thanks!
