FreakyJolly - ComfyUI VRAM Test Workflows (tested on an RTX 5080 16 GB, Oct 26, 2026)
Article: https://freakyjolly.com/how-much-vram-for-comfyui-16gb-test/

FILES
  01_Z-Image-Turbo_FP8.json        UI workflow (drag into ComfyUI)  - 1024x1024, 8 steps, res_multistep, cfg 1
  01_Z-Image-Turbo_FP8_API.json    same graph in API format (POST /prompt)
  02_Krea-2-Turbo_FP8.json         UI workflow - 1024x1024, 8 steps, euler/simple, cfg 1 (official template settings)
  02_Krea-2-Turbo_FP8_API.json     same graph in API format
  Only core ComfyUI nodes are used. No custom nodes needed. Tested on ComfyUI 0.38.1.
  To test VRAM at other sizes, change width/height/batch_size on the EmptySD3LatentImage node.

MODELS (put them in ComfyUI/models/...)
 Z-Image Turbo FP8:
  diffusion_models/z-image-turbo_fp8_scaled_e4m3fn_KJ.safetensors
     https://huggingface.co/Kijai/Z-Image_comfy_fp8_scaled
  text_encoders/qwen_3_4b.safetensors
     https://huggingface.co/Comfy-Org/z_image_turbo/tree/main/split_files/text_encoders
  vae/ae.safetensors
     https://huggingface.co/Comfy-Org/z_image_turbo/tree/main/split_files/vae
 Krea 2 Turbo FP8:
  diffusion_models/krea2_turbo_fp8_scaled.safetensors
  text_encoders/qwen3vl_4b_fp8_scaled.safetensors   (CLIPLoader type: krea2)
  vae/qwen_image_vae.safetensors
     all three: https://huggingface.co/Comfy-Org/Krea-2

MY RESULTS (RTX 5080 16 GB, Windows 11, 128 GB RAM, ComfyUI budget about 11 GB for weights)
  Z-Image Turbo FP8 1024x1024: 16.6 s per image warm, 0.57 s/it, fully loaded
  Krea 2 Turbo FP8 1024x1024: 26 min 21 s, 1.64 GB of weights offloaded, about 192 s/it
  Full table and explanation in the article.
