Easy Workflow - Krea2 + SesquiLSR (High Quality Latent Upscale) - 8GB Friendly
Two-stage text-to-image graph for Krea 2 in ComfyUI. (Image-to-image works fine)
Stage 1 generates at ~1 MP. Stage 2 upscales the latent with SesquiLSR — the learned upscaler that matches Krea 2’s Qwen / Wan VAE — then a light denoise refine. That is cleaner than bislerp and cheaper than decode → pixel upscale → encode.
Good settings to start with
- Stage 1: 8 steps, CFG 1, euler / simple (Krea 2 Turbo)
- Stage 2 SesquiLSR: scale 1.5, 8 steps, CFG 1, denoise 0.25
(Stage 2 denoise 0.25 is the Sesqui default. If it drifts, drop to 0.20. If it stays soft, try 0.30, even going as low as 0.12 can work great if you want to match original image but with better details.)
But of course... always fool around and find your own sweet spots!
Tested on RTX 4060 8 GB with ComfyUI --lowvram, outputting 2k and 4k images with latent upscale.
Custom nodes
1. SesquiLSR (required)
cd ComfyUI/custom_nodes git clone https://github.com/LoganBooker/SesquiLSR.git
Wan 2.1 weights upscaler_Wan21.safetensors) ship in that repo.
How to run
1. Update ComfyUI, install SesquiLSR, restart.
2. Drag this JSON workflow file onto the canvas.
3. Select your Krea 2 UNET.
4. Queue.