A Krea 2 text-to-image workflow for ComfyUI, the one I have been playing with. One prompt box, a LoRA stack, and four switches that decide how much work happens after the base render. One PNG per run, with the generation data and the workflow inside.

I benchmarked every switch: what each one costs and what it changes is in the comparison article. The short version is at the bottom of this page.

The four switches

• Use LoRAs: on, the model and the text encoder go through the LoRA stack, and the LoRA names go into the saved prompt for the metadata. Off, both skip the stack.
• Use bf16 Text Encoder: on, the bf16 Qwen3-VL 4B encoder. Off, the int8 one, about half the size. Take int8 when VRAM is tight.
• Use Hires Fix: on, the base picture goes through 4x-UltraSharp to 1.5x, then 6 more steps at denoise 0.35, with tiled VAE encode and decode. 1200x1600 comes out 1800x2400.
• Use Face Detailer: on, FaceDetailer finds faces with face_yolov8m and redraws them, 8 steps at denoise 0.35. It works on the hires picture when the hires fix is on, and on the base picture when it is off.

In the file all four are on, so you get the full stack.

Defaults

• 1200x1600 from CR Aspect Ratio (custom size). 10 steps, CFG 1, er_sde, simple.
• No negative prompt box. At CFG 1 the negative has no effect, so the graph feeds the sampler an empty one.
• One Seed node feeds the base pass, the hires pass and the face detailer.
• The checkpoint loads through Image Saver's UNet loader, so the model name lands in the metadata.
• Image Saver writes the prompt, LoRA names, seed, steps, CFG, sampler and scheduler into the PNG and embeds the workflow. The file name is time, model name and seed.
• A notes node holds the trigger words of my three character LoRAs (Shawy, Zara, Aria) and my house style sentence. It is wired to nothing. Copy from it.

LoRA slots

The stack ships with my Shawy LoRA at 0.7 in slot 1, plus two style LoRAs from my stack at 0.4 and 0.3. Swap in your own. My files sit in a subfolder, so open each slot once and pick your file.

Gotcha: set any LoRA you do not have to None. ComfyUI checks every dropdown before it runs, so a missing file stops the queue even with Use LoRAs off.

Required custom nodes

• Comfyroll Studio: Image Size, the LoRA stack, Apply LoRAs, the metadata prompt join, the 1.5x upscale.
• ComfyUI Image Saver, 1.23.0 or newer: the model loader, steps, CFG, sampler, scheduler, metadata and Save Image.
• comfy-image-saver by giriss: the Seed node only. A different pack from the one above, with almost the same name.
• Impact Pack: Face Detailer.
• Impact Subpack: the face detector.
• WAS Node Suite, the maintained copy: Text Multiline, the prompt box.
• Efficiency Nodes: LoRA Stack to String, the LoRA names for the metadata.

The switches, the booleans and the notes node are core ComfyUI. If they show red, update ComfyUI. Tested on ComfyUI 0.44.0.

Model files

• Checkpoint: RedCraft 3.0 (Krea 2), the int8 build, redcraft23_30Krea2_int8_convrot.safetensors, in models/diffusion_models.
• Text encoder: qwen3vl_4b_bf16.safetensors or qwen3vl_4b_int8_convrot.safetensors, in models/text_encoders, loader type krea2. Both loaders must point at a file that exists. With one encoder only, pick it in both.
• VAE: qwen_image_vae.safetensors, in models/vae.
• For the hires fix: 4x-UltraSharp.pth from OpenModelDB, in models/upscale_models.
• For the face detailer: face_yolov8m.pt from Bingsu/adetailer, in models/ultralytics/bbox.

What the benchmark says

From the comparison article, 89 renders on one RX 7900 XTX:

• At feed size the base picture is hard to tell from the full stack. On my card the base took about 36 s, everything on about 177 s.
• The hires fix pays when a picture is viewed big or has small details. In the test it turned a garbled pendant into a clean amber stone.
• To save about 70 s per hires render, swap the 4x-UltraSharp node for core Upscale Image By, set to lanczos and scale_by 1.5. At viewing size it looked the same, a little softer on photo hair. Two cases, one seed each.
• The face detailer at 0.35 changed my LoRA characters' faces: older on the photo look, red eyes toward brown, mottled cel shading. For LoRA characters, leave it off. Lower denoise I did not test.

Known limits

• Tested on one AMD card under ROCm. Nothing in the graph is AMD-specific.
• Tested on RedCraft 3.0 only, with two character LoRAs in three styles.

You are free to try it. Tell me in the comments if it works for you. If it breaks, paste the error text.