# 🔥 ONE MEMBERSHIP = FULL ACCESS

Your membership includes :

• All models

• Exclusive LoRAs

• Exclusive workflows

• Scripts & tools

• All future updates included

👉 $25/month on Ko-Fi

⭐ Get every current model, every future release, and all exclusive content with a single membership.

📢 Join The Community

We're building a friendly space to chat, share creations, and get support.

👉 Click here to join the Discord

KREA2FUSION - Character Fusion Workflow by Enzino

Swap the person in any photo with your own character LoRA, no prompt writing needed.

Load a source image, pick your character LoRA, type its trigger word, hit Run. The workflow does the rest: it analyzes the image, rebuilds the prompt automatically, and regenerates the picture with your character in place of the original person, keeping the scene, outfit, pose and lighting intact.

How it works

1. Auto Caption (Qwen3-VL) - The source image is captioned locally by Qwen3-VL 4B. The instruction is carefully designed to describe the scene, clothing, pose, lighting and background but never the face, hair or identity. This is the secret sauce: since the prompt says nothing about the original person's appearance, your character LoRA has full control over the identity.

2. Prompt assembly - Your LoRA trigger word (typed once in the LORA Prompt node) is joined with the auto-generated caption and fed into the sampler. A Prompt Preview node shows you the exact prompt used on every run.

3. Pass 1 - Identity swap - Img2img pass at denoise ~0.50: high enough for the LoRA to impose the new identity, low enough to preserve the composition.

4. Pixel-space upscale - The result is upscaled with 4xNomos2_hq_mosr (then scaled back to 2x). Real pixel upscaling instead of latent upscaling = much sharper output.

5. Pass 2 — Refine - A light second sampling pass (denoise ~0.30) cleans and sharpens the upscaled image.

6. Face Detail — A final FaceDetailer pass polishes the character's face.

Requirements

Models

• Krea 2 checkpoint (tested with IntoRealism Krea 2)

• qwen3vl_4b text encoder + wan_2.1_vae (standard Krea 2 setup)

• 4xNomos2_hq_mosr upscale model

• A character LoRA trained for Krea 2

Custom nodes (all available in ComfyUI Manager — use "Install Missing Custom Nodes")

• ComfyUI-QwenVL (by 1038lab) — the Qwen3-VL 4B caption model downloads automatically on first run

• rgthree-comfy

• ComfyUI Impact Pack (FaceDetailer)

First-run setup

Open the QwenVL — Auto Caption node and set it once:

• custom_prompt: already filled — this is the "describe everything except the face" instruction. Don't remove it.

• max_tokens: 512

• keep_model_loaded: off (frees VRAM before sampling — recommended on 16 GB cards)

• quantization: 4-bit on 16 GB VRAM or less

Then type your LoRA trigger word in the LORA Prompt node and load your LoRA in the Power Lora Loader.

Tuning tips

• Character doesn't take over enough? Raise Pass 1 denoise toward 0.55–0.60.

• Scene drifts too far from the source? Lower Pass 1 denoise toward 0.35–0.40 — the FaceDetailer will still lock in the identity.

• Check the Prompt Preview after each run. If it looks empty or weird, the caption node isn't configured - revisit the first-run setup above.

Runs comfortably on 16 GB VRAM.