✦ Enzino Krea2 Fusion v1.0.0

Same photo. Your character.ComfyUI nodes by Enzino, creator of IntoRealism

Drop a photo, pick a LoRA, run. The person becomes your character with the same pose, outfit and light, and the background stays pixel-identical.

Background locked ⚡ 4-seed preview Local caption · no API One-click model downloads
Quick start The four modes

The four modes

Fusion starts from a real photo, so the composition, pose and light are already right. You only choose what changes. In each diagram, amber is what gets redrawn and green is what stays identical.

Whole persondefault

With a character LoRA, the person becomes your character. Without a LoRA, they get a fresh look in the same scene.

Keep face

The original face and hair are protected. The LoRA reshapes the body, outfit and style. Works with body, height or age LoRAs.

Face only

Only the face becomes your LoRA, redrawn at 1024 px from the full-resolution photo. Everything else stays pixel-identical.

Identical backgrounddefault

Combined with Whole person or Keep face. Only the person's silhouette is redrawn: walls, logos, colours and sharpness stay identical.

Install

  1. Update ComfyUIKrea 2 and the local caption need a recent build. Fusion is tested on ComfyUI 0.39.
  2. Copy the folderPut ComfyUI-Enzino-Krea2-Fusion in ComfyUI/custom_nodes/.
  3. Install the one dependencyThe face detector and the person segmenter use ultralytics. It is already there if you use Impact Pack. 
    # standard install
pip install ultralytics

# ComfyUI portable (Windows)
python_embeded\python.exe -m pip install ultralytics
  4. Restart and open the workflowRestart ComfyUI, reload the page with Ctrl + F5, then open workflows/KREA2FUSION_Enzino.json.

No other node pack is needed: no Impact Pack, no rgthree, no external captioner. The only exception is ComfyUI-GGUF, if you want to load a .gguf model.

Models

The ① Models card shows an orange ⬇ button for every missing file. It downloads the file into the right folder and selects it.

FileFolderHow to get it
IntoRealismKrea2FlashV1.safetensorsmodels/diffusion_modelsCivitai, free
qwen3vl_4b_bf16.safetensorsmodels/text_encoders⬇ in ① Models · 8.9 GB
wan_2.1_vae.safetensorsmodels/vae⬇ in ① Models · 254 MB
4xNomos2_hq_mosr.pthmodels/upscale_models⬇ in ① Models · 17 MB
face_yolov8m.ptmodels/ultralytics/bbox⬇ in ① Models · 52 MB
person_yolov8m-seg.ptmodels/ultralytics/segm⬇ in ① Models · 55 MB
  • Any Krea 2 model works: IntoRealism Krea 2, Krea 2 Turbo, or a .gguf with ComfyUI-GGUF.
  • Equivalents you already have are recognised and picked automatically: qwen3vl_4b_fp8_scaled, qwen_image_vae, the .safetensors Nomos, and person_yolov8s-seg or n-seg.
  • The face detector and person segmenter are the same files as Impact Pack and ADetailer.

Quick start

  1. Drop a photo on the ③ cardDrag it in, click Choose…, paste with Ctrl + V, or pick one from Recent.
  2. Pick a LoRA in ②Click Add LoRA and type a few letters of its name. Its trigger word fills in by itself. With no LoRA, Fusion restyles the person.
  3. Choose the modeWhole person, Keep face or Face only, with Background: Identical.
  4. ⚡ Preview, then renderTry 4 seeds in a few seconds, click the best one, then Render this one in full.

Preview, render, export

The Krea2 Fusion card after a quick preview: settings on the left, the before and after slider in the middle, four seed previews on the right with the third one selected and an orange button to render it in full.
After ⚡ Preview ×4: the selected seed is outlined, and one click renders it in full. The pictures here are illustrations.
  1. ⚡ Preview ×NIn the Output header. It runs the first pass only, at the working size, on 2, 4, 6 or 8 seeds (set in Render). Nothing is saved.
  2. ▶ Render this one in fullLocks the chosen seed and runs the upscale and refine. Pass 1 comes from the cache, so the image you picked is exactly the one you get, only larger and sharper.
  3. ⬌ Before / afterIn the Output tools. It saves a side-by-side JPG at full height to output/Fusion/compare and downloads it, ready to post. The small enzinoai.com mark can be turned off in Output.

Click the output, or ⤢, for the full-screen viewer: wheel to zoom, drag to move, B before/after, C compare, ← → variations, ↗ new tab, ⬇ download.

The three cards

① Models

The Models card: Krea 2 model, text encoder and VAE under Generation; upscaler, face detector and person segmenter under Helpers.

Generation: the Krea 2 model, the text encoder and the VAE. The text encoder also writes the caption, so no extra caption model sits in VRAM.

Helpers: the upscaler, the face detector (Keep face, Face only) and the person segmenter (Identical background).

Advanced: weight_dtype and shift, 3 by default, as in the original KREA2FUSION graph.

A section opens by itself when one of its files is missing.

② LoRAs

The LoRAs card with the search open: typing krea2 lists every matching LoRA with its folder and trigger word, the matching letters highlighted.

Click a LoRA name to search the whole models/loras folder by name, sub-folder or trigger word. Your recent picks come first. Use ↑ ↓ and Enter to choose and Esc to close.

Each LoRA has a strength (−2 to 2), an on/off switch and a trigger word. The trigger word is read from .civitai.info, a .txt or the file's training metadata, and you can type your own.

Thumbnails: put my_lora.preview.png (or .png, .jpg, .webp) next to the LoRA file.

Several LoRAs stack, for example a character LoRA plus a height or body LoRA.

③ Krea2 Fusion

The Krea2 Fusion card after a full render: settings, the before and after slider, and the 2× output with its size and seed.
Settings, the before/after slider and the output side by side. The pictures here are illustrations.

Six sections, all collapsible: Source image, Transformation (mode, background, strength), Prompt (auto caption or manual, extra words), Render (preview count, working size, steps, seed), Finish (upscale, colours, grain) and Output (variations, saving, export mark). The header of each one sums up its current values. The EN button switches the interface to French.

Settings

The defaults are the original KREA2FUSION pipeline (pass 50 % · 8 steps · shift 3 → 4xNomos 2× → refine 30 % · 3 steps), with the background locked.

SettingWhat it does
TransformationHow much changes. 0 % unchanged · 10–25 % light touch, good for new characters from a face without a LoRA · 35 % subtle · 50 % balanced · 65 % strong. 50 %
BackgroundIdentical: only the person is redrawn and the background stays pixel-identical; needs the person segmenter, with a ⬇ button right under the option if it is missing. Similar: the background is lightly reworked with the person, same scene with small changes in details, text or logos. Identical
Face passesFace only: 2–3 lighter passes pull the face closer to the LoRA. 1
Main face · Every faceFace only: replace the largest face, or all of them. Main face
CaptionAuto: written locally by the Krea 2 encoder. It describes the scene but never the face, so your LoRA owns the identity, and it is cached per image. Manual: your own text. The optional prompt_in input takes any captioner. Auto
Add to the promptWords added after the caption, such as an outfit, an expression or the light.
Quick previewSeeds tried by ⚡ Preview: 2, 4, 6 or 8. 4
Working sizeLong edge of the first pass: 768, 1024, 1280, 1536 or Custom (256–2560 px). The card shows the exact working and final sizes. 1024
StepsFirst-pass steps. 8
SeedNew seed each run, or locked. new
Upscale + refineOff, 1.5× or 2× with the upscale model, then a light refine. 2× · 30 % · 3 steps
Keep source colors30–60 % if the LoRA shifts the colours. off
Film grain15–30 % breaks the plastic AI smoothness. off
VariationsSeveral images per run. The caption is computed once for all of them. 1
Save to output · File nameSaved to output/Fusion/. on · Fusion/Enzino

Tips

  • The LoRA doesn't take over: raise Transformation to 60–65 %, and check the trigger word.
  • The scene drifts (pose, outfit): lower Transformation to 35–45 % and keep Identical background.
  • New characters without a LoRA: Whole person at 10–25 % changes the face slightly while keeping the photo.
  • Face only for the cleanest result: body and scene are untouched, and no caption is needed, so it is the fastest mode.
  • Stay at 1024–1280 px for the first pass and let the 2× upscale do the rest. Above 1536 px, Krea 2 gets slow and can drift.
  • Every stage is cached: change only the upscale and pass 1 is reused, and the full render of a preview reuses its first pass.

Troubleshooting

"The caption came back empty"

The text encoder can't caption. Use qwen3vl_4b_bf16 or qwen3vl_4b_fp8_scaled, or switch Caption to Manual.

"background not locked" under the output

The person segmenter is missing or ultralytics isn't installed. Download the segmenter in ① Models → Helpers, install ultralytics in ComfyUI's own Python, then restart.

If the message says no person found, the photo has no clearly visible person: the whole image is redrawn.

A model has a red ring

The file saved in the workflow isn't on your disk. Pick another one, or click its ⬇ button in ① Models.

"No Krea 2 model found"

Put a Krea 2 model in models/diffusion_models, then press R in ComfyUI to refresh the lists.

Keep face or Face only does nothing

No face was detected: the note under the output says so. Check that the face detector is selected, and lower the detection threshold in Face only if the face is small or turned away.

The cards don't appear, only plain widgets

Reload the page with Ctrl + F5. If it persists, check the ComfyUI console for lines starting with [Enzino Fusion].

Out of memory

Lower the working size to 1024, use qwen3vl_4b_fp8_scaled, or set weight_dtype to fp8 in ① Models → Advanced. The upscale tiles automatically and retries with smaller tiles.