Gladas' Workflow for Krea 2 Turbo — v1 — v9

This Krea 2 Turbo v9 workflow creates images from text prompts and refines images with upscaling and detailers; it takes prompt text or an image and produces an image.

Images can go through two USDU passes; the second includes a before-and-after comparer and independent sampler, step, denoise, and LoRA controls.

VAE encode/decode stages sit before USDU 2 and each detailer under a shared VAE switch; fallback image switches follow USDU 2 and each detailer.

Wildcard text can be combined with other prompt text, and empty wildcard text remains empty.

Saved images can include metadata in a CivitAI-readable format.

The listed node packs are bkwildcards, ComfyMath, comfyui_ultimatesdupscale, comfyui-custom-scripts, comfyui-image-saver, comfyui-impact-pack, comfyui-impact-subpack, comfyui-kjnodes, and rgthree-comfy.

Listed model files include Anzhc Face seg 1024 v2 y8n.pt, bbox/Eyeful_v2-Individual.pt, betterDaysIllustriousXL_illuXLV02BAKEDVAE.safetensors, qwen_image_vae.safetensors, and qwen3vl_4b_fp8_scaled.safetensors.

The list also includes sam_vit_b_01ec64.pth, seedvr2_3b_int8_convrot.safetensors, seedvr2_ema_vae_fp16.safetensors, and vxpKrea2Nsfw_beta3INT8.safetensors.

Both USDU passes use the incoming image's width and height for tile dimensions; disconnect those inputs to set smaller tiles manually.

Initial generation and both USDU schedules use Normal; the detailers use Simple.

The first USDU pass uses 3 steps at 0.22 denoise, the second uses 2 steps at 0.15, and both detailers default to 2 steps at 0.15 denoise.

Tip: compare results with the optional second USDU pass and bypass it when it does not help; it may improve or worsen an image.

Suggestion · non vérifié

The workflow is not set up for models using CFG above 1; full-image tiles can increase memory use and processing time, depending on model and hardware.