import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("krea/Krea-2-Raw", dtype=torch.bfloat16, device_map="cuda")
pipe.load_lora_weights("conradlocke/krea2-identity-edit")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]Krea 2 Identity Edit β V2
β Support on Ko-fi β all tips go straight to GPU compute for future versions.
π§° Training code is public: krea2edit-trainer β the ai-toolkit extension these LoRAs were trained with, geometry-matched to the nodes, with measured consumer-GPU VRAM requirements.
V2 ships as two LoRAs:
krea2_identity_edit_v2.safetensorsβ Krea 2 Identity Edit V2. Put a person from one photo into any scene, pose, outfit and lighting. Extremely improved likeness, better sharpness and skin/fabric texture, and far more freedom to restage than v1.2.krea2_general_edit_v2.safetensorsβ Krea 2 General Edit V2. Instruction-based editing: remove, recolor, relight, swap backgrounds, add objects, change text and outfits, restyle, flat lays.License: free for personal / non-commercial use; commercial use needs a license from me, case by case β see License.
Use one LoRA per task β they share a lineage, so loading both at full strength breaks the image (see Using both).
Requires the V2 nodes (
comfyui-krea2editβ₯ 2.0.0), which add reference placement: choose per reference whether it is pinned onto the output layout or placed beside it β e.g. keep a scene exactly while placing a person freely in it.Low-VRAM versions:
..._v2_r128.safetensors(0.91GB) and..._v2_r64.safetensors(0.46GB) for both LoRAs β rank-reduced from the full files, near-identical quality.Prior versions remain available (
..._v1_2*,..._v1_1*,..._v1).
Identity-preserving restaging and instruction-based editing for Krea 2 (12.9B single-stream MMDiT). Give it an image and a plain-language instruction.
An unofficial community fine-tune of Krea 2 Raw. Not an official Krea product; not affiliated with or endorsed by Krea.ai, Inc.
Requires the ComfyUI-Krea2Edit node pack β the LoRAs are trained with dual conditioning (in-context VAE tokens + image-grounded Qwen3-VL encoding) that stock nodes don't provide.
Krea 2 Identity Edit V2
One photo of a person in, that person anywhere out:
- Restaging: new scenes, poses, camera angles, eras, seasons, from behind, full body or close-up.
- Lighting: neon, direct flash, golden hour, stage spots, candlelight, underwater β the light actually falls on the face.
- Wardrobe: describe the outfit and the model dresses the person in it.
- Compositing: add a person into an existing scene, or several people one after another into a group photo.
- Character sheets: make a turnaround sheet from a photo, or a photo from a sheet.
Recommended settings (Identity Edit V2)
| Use | Model | Steps | CFG | ref_boost |
grounding_px |
Placement |
|---|---|---|---|---|---|---|
| Restage one person (single reference) | Turbo | 8 | 1.0 | 2 | 768 | beside or aligned |
| Add a person to a scene (scene = image 1, person = image 2) | Turbo | 8 | 1.0 | person 2β3, scene 1 | 768 | scene aligned, person beside |
| Face-locked portraits (keep the framing of the photo) | Turbo | 8 | 1.0 | 1β2 | 768 | aligned |
Prompting matters more than the dials:
- Say how much of the person is visible. "Standing on a ridge" is ambiguous and tends to copy the photo's head-and-shoulders crop; "in muddy leather boots on a ridge", "sitting cross-legged on a log", "full body" gets the whole person.
- Describe the outfit concretely, and what should be gone. The model leans on the clothing in the reference photo. "Wearing only a white tank top, bare shoulders, visible collarbones" or "remove her coat" works far better than just naming a new garment.
- Describe light by where it lands: "a magenta sign paints the left half of her face pink" beats "neon lighting".
Krea 2 General Edit V2
Superior image editing from a plain instruction: remove objects and people, recolor, relight, change the season, swap backgrounds, add objects, rewrite sign text, change outfits, restyle (e.g. anime), and turn an outfit into a product flat lay β while keeping everything you didn't ask to change.
Recommended settings (General Edit V2)
| Task | Model | Steps | CFG | ref_boost |
grounding_px |
|---|---|---|---|---|---|
| Most edits | Turbo | 8 | 1.0 | 4 | 768 |
| Removals / large deletions | Raw | 20 | 3.0 | 4 | 768 |
At CFG > 1, ground the negative too (empty prompt + same image).
Reference placement (new in the V2 nodes)
The model sees each reference image as a grid of tokens next to the image being generated. Two new node inputs choose where each reference sits:
ref_positionβ the last reference (the person in two-image workflows, or the only image)ref_position_aβ the first reference (the scene)
Options:
aligned (trained)(default) β the reference is pinned onto the output layout. Best when the layout should stay: edits, face-locked portraits, the scene in a composite.beside (right of target)β the reference sits next to the canvas: identity still flows, but the layout pin is gone, so the person can be restaged more freely. Scene aligned + person beside gives the most natural composites.bag of featuresβ experimental, all tokens of a reference share one position (content without layout). No recommended use yet.
The models were trained with aligned only; beside works well in our testing but
is new.
Using both LoRAs together
Both LoRAs build on the same lineage, so loading both at strength 1.0 produces pure noise. If you want them combined, use about 0.6 each; in our tests that gives the identity model's likeness together with the general model's removals and recolors. Otherwise, use one LoRA per task.
Known limitations (honest list)
- Identity Edit V2 is not a general editor. It tends to ignore removals, recolors and camera changes β use General Edit V2 for those.
- Clothing from the reference photo can carry over (coats, turtlenecks). Name the new outfit concretely and say what is gone (see prompting above).
- Objects from the reference can leak into a restaged scene occasionally (e.g. a candle from the source photo's table). Reroll or describe the scene more fully.
- Head swaps tend to become whole-person swaps (the swapped person's clothes come along).
- Low-resolution face restoration and outpainting are not reliable.
- Large groups: adding people one at a time works well up to about seven; beyond that the frame fills and people start to replace each other.
- Generate at β€2MP.
License
V2 models: free for personal and non-commercial use. Commercial use needs a license from me, agreed case by case β contact l.bouaraba@gmail.com.
krea2_identity_edit_v2.safetensorsandkrea2_general_edit_v2.safetensorsare distributed under the Krea 2 Identity Edit V2 License and the Krea 2 Community License Agreement (see alsoNOTICE).- Free: personal projects, hobby art, research, education, evaluation, sharing your images unpaid.
- Needs a commercial license: selling or licensing outputs, client/freelance work, ads and marketing, products, offering the model or a service built on it, and any use by or for a company. Merges and fine-tunes of V2 carry the same terms.
- Krea's terms apply on top: deployments must implement reasonable content moderation (Β§4.2), AI disclosure obligations apply where required (Β§4.3), and entities with $1M+ annual revenue also need an enterprise license from Krea (Β§2.3).
- Earlier versions (v1, v1.1, v1.2) stay under the Krea 2 Community License Agreement as released.
This repository modifies the Krea Model as permitted by Β§3; it is not endorsed by Krea. Research/portfolio release by Lars Bouaraba, a self-funded hobbyist.
Showcase
All people shown are themselves AI-generated β no real likenesses.
V2 β Identity Edit
V2 β General Edit
Prompts are printed under each image.
v1.1
v1 gallery
Support
This is a solo hobby project, no company or funding behind it, and it stays free and open either way. If it's useful to you and you'd like to chip in toward GPU compute for future versions, there's a tip jar: ko-fi.com/conradlocke. Completely optional. Using the model, sharing results, and reporting issues already helps a lot. Thank you.
Contributors & thanks
This is a solo project, made a lot better by the community. Thank you to:
- stablellama β the MIT-licensed head/face/eye/person swap dataset behind those capabilities since v1.2.
- CeciliaXCIX β tireless, high-quality community support in the discussions.
- akashzeno β node engineering: diagnosing the ComfyUI compatibility break and contributing the regression test.
- ethanfel β the
target_latentpre-encode contribution to the nodes. - SubtleShader β testing the training code and consumer-GPU feedback.
- Mark (sogni.ai) β support, a GPU-fund donation, and getting the word out.
Want to help? Contributions of training data and node/code work are welcome, see the discussions.
Scope and responsible use
Krea 2 Identity Edit is an identity-preserving character restaging model, trained only on SFW data. It is not trained on any NSFW concepts, and I have no plans to add or support NSFW data in current or future versions.
I do not endorse or support using this model to produce non-consensual, harmful, or sexual imagery of real people, including deepfakes. Please use it responsibly and respect the consent and likeness of anyone you depict.
- Downloads last month
- 153,255