Model weight and version guidance
Prefer v1.2 when it is available. It is the recommended build on the model card and adds reference sheets, swaps, inpainting, outpainting, virtual try-on, and FIT geometry. Confirm the version pinned by the hosted Space before comparing results across sessions.
Recommended file
krea2_identity_edit_v1_2.safetensorsSettings that change Krea2 Edit results
Treat these values as starting points from the public model card and node changelog, not universal guarantees. Change one control at a time and compare against the source image.
| Job type | Checkpoint | Steps | CFG |
|---|---|---|---|
| Add / recolor / restyle / re-stage | Krea 2 Turbo | 8–12 | 1.0 |
| Removals / large deletions | Krea 2 Raw | ~20 | 3.0 |
- ref_boost ≈ 4 — strong likeness starting point on v1.2 nodes; above ~10 starts breaking removals; below 1 loosens creativity.
- grounding_px — lower follows the edit more uniformly; higher protects identity. Duplicate “split picture” frames usually mean the value is far above the trained range (v1.1 docs cite 384–768 as the comfort band).
- LoRA strength 1.0 — published default. At CFG above 1, ground the negative with an empty prompt plus the same image.
- Low-VRAM SVD cuts of v1.2 (
r128~ 0.91 GB,r64~ 0.46 GB) retain >99% of weight energy per the card — useful when you leave browser Krea2 Edit for a laptop GPU.
Browser Krea2 Edit vs local ComfyUI
The hosted editor is best for quick prompt tests. A local graph is better when you need exact versions, repeatable seeds, batches, private infrastructure, or additional LoRAs.
Hosted Space (this page)
- No checkpoint download; try Krea2 Edit in minutes
- Good for prompt R&D and client demos on Krea2 Edit
- Depends on Space uptime and shared GPU queues
- Open full-screen: hf.space
Local ComfyUI-Krea2Edit
- Pin v1.2 weights and FIT / ref_boost node versions
- Stack character, body, or style LoRAs on top
- Needs native Krea 2 + Qwen3-VL encoder support
- ComfyUI-Krea2Edit on GitHub
Limits, license, and responsible use of Krea2 Edit
Likeness on Krea2 Edit is described upstream as texture-faithful and proportion-conservative: skin and lighting often transfer well, while strongly unusual facial geometry may read like a close relative. Outfit swaps are hit-or-miss; removals need the Raw recipe and still fail sometimes. Extreme hair or body types may need a subject LoRA on top.
Weights are a Derivative Model under the Krea 2 Community License. Commercial use is allowed under the published revenue threshold (model-card summary: under $1M/year); above that, contact Krea for enterprise terms. Deployments should moderate content; AI disclosure may apply. The LoRA is SFW-trained — do not create non-consensual or harmful likenesses of real people. Read our Privacy Policy and Terms of Service.
Technical sources
Check the upstream model card and repositories for current files, workflows, changelogs, licenses, and compatibility notes.