Krea 2 Character Consistency in ComfyUI: Full Workflow Guide

Getting the same character to show up correctly across ten different images used to mean training a LoRA on that character first. Krea 2 changes this — you can hand it one reference photo and a text instruction, and it generates a new pose or scene while keeping the face, outfit, and hairstyle close to the source. No training required.

This guide walks through the exact workflow: which nodes to connect, what each setting does, and how to write instructions that keep identity intact instead of drifting into a different-looking person by the third generation.

If you're working with Qwen Image Edit instead of Krea 2 — including multi-character scenes with two or more people in one frame — see the general ComfyUI character consistency guide. This page focuses specifically on the Krea 2 Turbo node chain.

Tested on: ComfyUI v0.3.7+, RTX 4090 (24 GB VRAM).
Minimum VRAM: 12 GB (FP8 model).

How Identity Is Preserved

Two separate paths carry your reference photo into the sampler, and both matter:

• The text-encoder path — the reference image is fed into the Qwen3-VL text encoder alongside your instruction, so the encoder "sees" the image while it reads what you're asking for.
• The image-latent path — the same reference image is separately VAE-encoded into a latent and merged into the conditioning through ReferenceLatent. This is the path doing most of the work to hold the face steady.

Example: One Reference, Two Scenes

Below is one reference photo and two separate generations from that same reference — swap in your own screenshots here. Click any image to zoom.

🔍 Click to zoom

🪑 Same reference · 1280×1280 · steps 12 · cfg 2.5

🔍 Click to zoom

New Pose + Setting — Luxury Hotel Lounge

Copy Prompt▶View Instruction

🌆 Same reference · 1280×1280 · steps 12 · cfg 2.5

🔍 Click to zoom

New Pose + Setting — Second Run, Same Reference

Copy Prompt▶View Instruction

🌆 Same reference · 1280×1280 · steps 12 · cfg 2.5

🔍 Click to zoom

New Pose + Setting — Third Run, Same Reference

Copy Prompt▶View Instruction

What You Need

Krea 2 ships as separate model, text encoder, and VAE files — ComfyUI does not bundle them into one checkpoint. Download these before you start.

Where to Place the Files

Each file type goes in a different subfolder under ComfyUI's models/ directory. A file in the wrong folder means ComfyUI can't find it and the dropdown stays empty.

Download the Workflow

How to Load the Workflow JSON into ComfyUI

1. Download the .json file to your computer.
2. Open ComfyUI in your browser — usually at http://127.0.0.1:8188.
3. Drag the .json file directly onto the ComfyUI canvas. Drop it anywhere on the empty grey area.
4. The workflow loads automatically. You should see the model loaders, image loader, prompt, and sampling groups appear on the canvas.
5. If you see only red error blocks, jump to the Troubleshooting section below.

🔍 Click to zoom

Workflow Structure: What Each Node Does

Eleven nodes, connected left to right. Here's what each one controls:

🔍 Click to zoom

Step-by-Step Setup

1. In VAELoader, confirm Wan2_1_VAE_bf16.safetensors is selected.
2. In CLIPLoader, select qwen3vl_4b_fp8_scaled.safetensors and confirm type reads krea2.
3. In UNETLoader, select your Krea 2 Turbo model file.
4. Click LoadImage and upload your reference photo. Use a clear, well-lit, front-facing shot — side profiles and heavy shadows make identity lock harder.
5. Open the TextEncodeQwenImageEditPlus (Positive) node and write your instruction — see the next section for the exact structure.
6. Leave the negative prompt as-is (rejecting diptych/split-screen output) unless you see a different unwanted pattern.
7. Set your output width and height in the PrimitiveInt nodes — this workflow defaults to 1280×1280.
8. Click the orange Queue Prompt button. A progress bar appears below it. On an RTX 4090 at 12 steps, generation takes roughly 15–25 seconds.

How to Write the Instruction

Structure every instruction in two parts: what to preserve, then what to change. Lead with the preservation language — it carries more weight at the start of the prompt than buried at the end.

KEEP the same [subject] and change the pose and setting: [new pose, new location, new framing — be specific about posture, hands, expression, camera angle]. Preserve exact facial identity, makeup, hairstyle, and clothing details from the reference image.

Worked example from this workflow:

KEEP the same woman and change the pose and setting: she is now walking barefoot along the shoreline during a beautiful golden sunset, gentle ocean waves touching the sand around her feet. She looks over her shoulder toward the camera while one hand lightly brushes her hair as the sea breeze moves naturally around her. Relaxed, elegant posture with a peaceful, confident expression. Preserve exact facial identity, makeup, hairstyle, body proportions, clothing, accessories, and every visual detail from the reference image.

Negative Prompt

Krea 2 occasionally generates unprompted side-by-side comparison panels. The negative prompt in this workflow suppresses that pattern specifically:

diptych, split screen, two panels, side by side, before and after comparison, collage, multiple images

Settings That Matter

🔍 Click to zoom

Krea 2 Turbo is a distilled model built for low step counts. Going much higher than 12–16 steps doesn't meaningfully improve identity and just slows generation down. Pushing CFG higher forces more literal instruction-following but tends to flatten the natural look Krea 2 is known for.

Going Further: LoRA-Based Identity Preservation

If you need tighter identity lock than the base workflow gives you — especially for scene-referential instructions like "the woman on the left" — the community has released a dedicated LoRA for this. It's called Krea 2 Identity Edit, released about two weeks after Krea 2 itself went open-source, and it requires an additional node pack rather than the stock nodes covered above.

Need multiple characters in the same frame instead of one at a time? That's a different node setup — see the multi-reference section of the general ComfyUI character consistency guide.

"This node type does not exist" (red node in workflow)

The workflow uses a custom or newer core node you haven't installed.

1. Click Manager in the top menu.
2. Click Install Missing Custom Nodes.
3. Restart ComfyUI completely.

If nodes still show red, update ComfyUI itself — TextEncodeQwenImageEditPlus and ReferenceLatent are recent core additions and won't exist on older builds.

The character's face looks different in the output

This usually means the VAE-encode path isn't reaching the sampler correctly, or the reference photo is a poor angle for the model to lock onto.

1. Confirm VAEEncode is connected from your LoadImage node, not left disconnected.
2. Confirm ReferenceLatent receives both the positive conditioning and the VAE-encoded latent.
3. Try a clearer, more front-facing reference photo.

Out of memory / CUDA error during generation

The BF16 model is roughly 24.5 GB and can exceed VRAM on cards under 24 GB.

1. Switch to krea2_turbo_fp8_scaled.safetensors instead of the BF16 version.
2. Lower your output resolution (try 1024×1024 instead of 1280×1280).
3. Close other GPU-heavy applications before queuing.

Output comes out as a split-screen or diptych image

Krea 2 occasionally generates side-by-side comparison panels unprompted. Confirm the negative prompt still contains the diptych/split-screen rejection text — if it's been cleared or edited, restore it.

Frequently Asked Questions

Published: 2026-07-13 · Last updated: 2026-07-13 · Tested on RTX 4090 (24 GB VRAM) · ComfyUI v0.3.7 · Model: krea2_turbo_bf16.safetensors