Version 1 is a very simple txt2img Krea2 Workflow that does not have a upscale step. Version 2 upscales with SeedVR2 and then with ZImage which is what I found to give the best results. ZImage acts as a upscale fix based on my testing.
See this post about V3
https://civarchive.com/articles/36952/krea-2-on-12gb-vram-what-actually-improved-my-workflow-and-what-didnt

V3

For a first-time user:

• Use a current ComfyUI version with Krea 2 and native SeedVR2 support. Older installations may show missing nodes. Krea setup guide.

• Install the custom nodes: LoRA Manager, Image Saver, and ComfyUI-VOSR2.

• Download the models separately. The workflow JSON contains no weights. It needs the Krea checkpoint, Qwen text encoder, Sharp Plus VAE, SeedVR2 model/VAE, and VOSR2 bundle. Download the selected LoRAs or remove/disable their entries.

• Check every model dropdown after importing. Their local filenames may differ from yours.

• Expect substantial downloads and memory use. VOSR2 alone downloads roughly 7 GB on first use. Your 12 GB RTX 4070 runs the configured workflow, but that isn’t a tested minimum specification.

• The saved resolution is 1248×1824 → 2496×3648. Both downstream resize nodes use fixed dimensions; changing the base size alone won’t automatically change the final size.

V1