| File | krea2_turbo_fp8_scaled.safetensors |
| Model | Krea 2 (Turbo) (image model, 12.8B) |
| Precision | FP8 |
| Size | 13.1 GB |
| VRAM to fit | about 15.7 GB |
| Folder | ComfyUI/models/diffusion_models/ |
| Loader node | Load Diffusion Model (UNETLoader) |
| Source | Comfy-Org/Krea-2 on Hugging Face |
Download krea2_turbo_fp8_scaled.safetensors (13.1 GB)
Where to put it
Put the file in ComfyUI/models/diffusion_models/ (portable version: ComfyUI_windows_portable/ComfyUI/models/diffusion_models/). Older guides say models/unet: that folder still works. Press R in ComfyUI or restart it, then pick the file in Load Diffusion Model (UNETLoader). Not for Mac: FP8 and INT8 files do not load on Apple Silicon.
If a workflow says “Value not in list” for that node, the workflow expects a file with another name. Click the node and choose this one.
Which graphics cards it fits
| VRAM | This file | Best file for that card (RTX 40/50) | |
|---|---|---|---|
| 6 GB | Not practical | Short by 9.7 GB | krea2_turbo-Q2_K.gguf (4.9 GB) |
| 8 GB | Offload only | Short by 7.7 GB: runs with part of it in system RAM, slower | krea2_turbo-Q2_K.gguf (4.9 GB) |
| 10 GB | Offload only | Short by 5.7 GB: runs with part of it in system RAM, slower | krea2_turbo-Q3_K_M.gguf (6.0 GB) |
| 12 GB | Offload only | Short by 3.7 GB: runs with part of it in system RAM, slower | krea2_turbo-Q5_K_M.gguf (8.9 GB) |
| 16 GB | Runs well | Fits in VRAM, 0.3 GB to spare | this file |
| 20 GB | Runs well | Fits in VRAM, 4.3 GB to spare | this file |
| 24 GB | Runs well | Fits in VRAM, 8.3 GB to spare | this file |
| 32 GB | Runs well | Fits in VRAM, 16.3 GB to spare | krea2_turbo_bf16.safetensors (26.3 GB) |
| 48 GB | Runs well | Fits in VRAM, 32.3 GB to spare | krea2_turbo_bf16.safetensors (26.3 GB) |
Need = file × 1 + 1.8 GB for the picture and 0.8 GB for Windows and the browser — the same rule as the rest of the site. FP8 files do not load on a Mac (Apple Silicon); use a 16-bit or GGUF file there.
Measured with this exact file
| GPU | Time | Size | Steps | Peak VRAM | Source |
|---|---|---|---|---|---|
| RTX 5060 Ti 16 GB | 17.7 s per image | 1024x1024 | 8 | 14.9 GB | measured by me, 2026-10-04 |
Time = sampling + VAE decode, after a warm-up run. Peak VRAM = memory in use on the whole card. How I measure.
What else you need
This file alone does not make a picture. With it you load:
| Role | File | Size | Folder |
|---|---|---|---|
| Model | krea2_turbo_fp8_scaled.safetensors | 13.1 GB | models/diffusion_models |
| Text encoder | qwen3vl_4b_fp8_scaled.safetensors | 5.2 GB | models/text_encoders |
| VAE | qwen_image_vae.safetensors | 0.3 GB | models/vae |
18.6 GB in total. Smaller text-encoder versions and other options are on the Krea 2 page.
Other files of Krea 2
| File | Precision | Size | VRAM to fit |
|---|---|---|---|
| krea2_turbo_bf16.safetensors | 16-bit | 26.3 GB | 28.9 GB |
| krea2_turbo-Q8_0.gguf | Q8_0 | 13.7 GB | 16.3 GB |
| krea2_turbo_int8_convrot.safetensors | INT8 | 13.5 GB | 16.1 GB |
| krea2_turbo_fp8_scaled.safetensors | FP8 | 13.1 GB | 15.7 GB |
| krea2_turbo-Q6_K.gguf | Q6_K | 10.6 GB | 13.2 GB |
| krea2_turbo-Q5_K_M.gguf | Q5_K_M | 8.9 GB | 11.5 GB |
| krea2_turbo-Q4_K_M.gguf | Q4_K_M | 7.5 GB | 10.1 GB |
| krea2_turbo-Q3_K_M.gguf | Q3_K_M | 6.0 GB | 8.6 GB |
| krea2_turbo-Q2_K.gguf | Q2_K | 4.9 GB | 7.5 GB |
Same model, different compression. 8-bit files (FP8, INT8, Q8_0) look almost the same as 16-bit; below Q5 the loss starts to show.
Want the answer for your exact card? Check your GPU. Every file of every model: ComfyUI model files.
Size and source read from Hugging Face (2026-10-01).