A complete ComfyUI workflow for Krea 2 / Krea 2 Turbo that generates fast with Krea and then refines the image with an SDXL hires pass and per-character face detailers, so you get Krea's composition and prompt understanding with the accuracy of your SDXL character LoRAs.

## How it works

1. Krea 2 generation – natural-language prompt, 8 steps, CFG 1 (euler_ancestral / beta). Recommended resolution 832x1216 (2:3).

2. SDXL hires – upscale (remacri) to 1248x1824 and refine with an SDXL checkpoint + your LoRAs. The Krea prompt can be reused here; the LoRA carries the character.

3. Per-character face detailers – each face gets its own LoRA and its own short prompt/negative, so two characters in the same image don't blend into each other.

4. Speech-bubble protection – a speech-bubble segmentation model detects balloons and text, and the hires/detailer skips those areas, so lettering stays clean and readable. If no bubbles are found, the hires runs normally on the whole image.

## Features

- Organized prompt subgraph: Quality, Global Negative, and a separate Positive/Negative for each character

- Preview before hires, so you can discard bad gens early

- Works with or without speech bubbles

- Works for single characters or duos

## Tips

- Krea ignores the negative prompt at CFG 1, so describe what you want in positive terms. The negative only matters in the SDXL stages.

- Don't load both character LoRAs in the hires stage at full strength – it causes face fusion. Keep each character LoRA only in its own face detailer.

- For the face detailers, use short booru tags (trigger + key face traits), not the long Krea prompt.

- Put each character's traits in the other character's negative (e.g. eye color, eyebrows) to stop features from leaking.

## Requirements

- Krea 2 / Krea 2 Turbo (fp8_e4m3fn recommended), qwen3vl_4b text encoder, qwen_image VAE

- An SDXL checkpoint (any anime/furry SDXL model)

- 4x remacri upscaler

- Custom nodes: ComfyUI Impact Pack, Impact Subpack, rgthree (Power Lora Loader)

- Speech-bubble segmentation model (yolov8m_seg-speech-bubble) in models/ultralytics/segm

## Limitations

- Text outside of speech bubbles may still get refined.

- Complex multi-speaker balloons generated by Krea can be unreliable – adding the lettering afterwards gives the best results.

- Needs enough VRAM to keep the Krea and SDXL models loaded; the pipeline is slower than a single pass.