Krea 2 Two Characters Without Merging: Render Alone, Cut Out, Paste, Unify

Put two characters in one picture without sharing their hair, skin and clothes. The graph renders each character alone with their own prompt, cuts them out, pastes them on an empty scene where you choose the positions, and runs one low-denoise pass so the light matches. You also get exact control of who stands where and how big they are.

This graph needs a custom node: BRIA RMBG. Install ComfyUI-BRIA_AI-RMBG (nodes BRIA_RMBG_ModelLoader_Zho and BRIA_RMBG_Zho) and put the weights model.pth from briaai/RMBG-1.4 on Hugging Face in that node's RMBG-1.4 folder. BRIA publishes this model for non-commercial use, so check their terms before using the results commercially. Everything else is core ComfyUI.

REQUIRED MODELS

• Diffusion model: RedCraft 3.0 Krea 2, the INT8 convrot build, redcraft23_30Krea2_int8_convrot.safetensors, in models/diffusion_models. Plain Krea 2 and Krea 2 Turbo load too; RedCraft is what I tested.

• Text encoder: the Qwen3-VL 4B text encoder from the Krea 2 downloads (see the Krea 2 guide on comfyui-wiki.com); I run a community int8 ConvRot build of it (qwen3vl_4b_int8_convrot), and the bf16 or fp8 file works the same. It goes in models/text_encoders, loaded with CLIPLoader type krea2.

• VAE: the Qwen-Image VAE, qwen_image_vae.safetensors, in models/vae, from Comfy-Org on Hugging Face.

HOW IT WORKS

1. Character A and character B are each rendered alone in front of a plain grey backdrop, 896x1152, with their own prompt.

2. A third render draws the empty scene, 1152x896.

3. BRIA RMBG cuts out both characters. The mask is shrunk by 1 px and softened, so no grey halo.

4. Both cutouts are scaled by 0.62 and pasted on the scene. The two red nodes hold each cutout's x and y.

5. A pass at denoise 0.4 over the whole picture, with a prompt that describes the whole scene and both people, matches the light and the edges.

The graph saves the picture before the last pass and the final one, and previews each part.

HOW TO USE IT

• Write each character as a full figure, head to boots, ending with "the only person in the frame". Keep the plain grey backdrop line; the cutout depends on it.

• Write the scene prompt with nobody in it.

• Write the last prompt as the whole scene with both people, copying the two character descriptions. Say who stands left and right and which way the light falls.

• Set the positions. The defaults fit a 1152x896 scene.

• Use the lowest unify denoise that makes the join look like one picture.

SETTINGS

• Steps 8. It is a distilled model, more steps do not help. CFG 1.0, sampler er_sde, scheduler simple. Decode is VAEDecodeTiled, tile 512, overlap 64. Negatives: a ConditioningZeroOut, because a negative prompt does nothing at CFG 1.0.

• Unify denoise 0.4. I also ran 0.3 and 0.5 on one pair.

WHAT I TESTED, AND WHAT I FOUND

I tested with generic adults only: a woman in a yellow raincoat and a man in a navy cardigan, two women of different ages, two women with swapped colours (blonde in black, dark-haired in white), and two men with glasses and a beard against a man with a bun. The graph ran each pair from a clean start in about 52 seconds.

• Plain Krea 2 held these pairs well. In the single-prompt renders of the same pairs that I looked at, I did not see features swap between the two people. If your pairs merge, this graph is for you. If they do not, you do not need it for merging.

• What it did change: one thing my single prompt ignored was hair length. I asked for waist-length hair on both women of the colour-swap pair, and all eight single-prompt renders came out shoulder-length. Rendered alone, the women kept the long hair in both runs. You also get exact placement and size, which a single prompt cannot give.

LIMITS

• The last pass redraws faces a little. At 0.3 they stay closest to the cutout; at 0.5 faces and a beard were redrawn in my test. Clothes and hair colour held at 0.3, 0.4 and 0.5.

• Both characters stand apart. Hugging, holding hands or overlapping bodies need more than two cutouts; this graph does not do that.

• It needs a plain backdrop behind each character. A busy backdrop leaves bits in the cutout.

• I did not test character LoRAs in this graph. In my own work another person in a LoRA frame can pick up the LoRA's look, which is one reason to render each person alone, but I have not run that case here.

• Tested on RedCraft 3.0 (Krea 2) on ROCm (a 7900 XTX). The UI file was built by script and checked against the node definitions, not saved from a browser; if a node shows red on load, tell me which.

If this worked for you, a thumbs-up here helps the next person find it.