Krea 2 Outpaint — Any Side, One Pass (AnyPaint)

Extend a photo on any side in one Krea 2 Turbo pass. The AnyPaint LoRA and a model patch from my AusBoss nodes make Krea 2 continue the scene into the padding instead of painting a new one beside it, and Stitch Inpaint puts every original pixel back. The only custom pack you need is my free AusBoss nodes.

The first run extends the 16:9 photo from the sample inputs to a 4:5 portrait frame, top and bottom at once, at 1136 × 1424 on the 1.6 MP canvas.

What's inside

• Pad by dragging: Load Image + Pad. Drag any edge of the dashed rectangle outward; all four at once is fine. The padding is flat gray, the mask covers exactly the new area, and a 1.6 MP budget holds the whole canvas at a size Krea 2 paints cleanly, whatever the photo's size.

• A prompt that knows the photo: Krea 2's own Qwen3-VL text encoder describes the unpadded photo through the core Text Generate node, and your Outpaint direction line is added to it. Show Text shows the combined prompt before anything renders. No extra caption model.

• AnyPaint + model patch: the LoRA was trained with a gray-padded canvas as its reference. The Krea 2 Outpaint Model Patch, set to whole canvas, registers that reference over the full frame and caches its keys and values once per run.

• Known pixels stay put: the padded canvas is VAE-encoded with the pad mask (Set Latent Noise Mask), so the sampler only paints the gray.

• Seams: Stitch Inpaint lays the untouched source back over the result through a 64 px feather, and its tone match lifts the new area onto the source's tone along each seam.

• Compare slider, render timer, a PNG with the workflow embedded, and a Workflow Note card that lists every model with a download button and marks the files you already have.

Quick start

1. Update ComfyUI (I tested on 0.37.0).

2. Install AusBoss nodes from ComfyUI-Manager (search "AusBoss", registry id ausboss-nodes) and restart. You need 2.2.0 or newer, which is what Manager installs.

3. Unzip sample_inputs.zip into ComfyUI/input/ (it holds ausboss_day_bridge_169.png), or load your own photo in Load Image + Pad.

4. Drag the workflow in and download the four models from the card. If you keep models in subfolders, pick them in the loaders.

5. Queue. For your own photo, drag the pad edges and rewrite Outpaint direction to say what belongs in the new area.

Models

• krea2_turbo_fp8_scaled.safetensors → models/diffusion_models · 13.1 GB

• qwen3vl_4b_fp8_scaled.safetensors → models/text_encoders · 5.24 GB

• qwen_image_vae.safetensors → models/vae · 254 MB

• krea2_anypaint_rank32.safetensors → models/loras · 229 MB

Other Krea 2 builds (BF16, INT8, NVFP4) are on the same page: Comfy-Org/Krea-2. The loader is set to the official FP8 Turbo file; my test runs used a different local FP8 Turbo build.

Custom nodes

Only ComfyUI-AusBoss, 2.2.0 or newer (Manager installs it): free and open source. Everything else is core ComfyUI, including Text Generate, Set Latent Noise Mask and the sampler. You don't need the AnyPaint node pack, an LLM pack or rgthree.

Manager can't find AusBoss? In the classic Manager window, if the Channel box is empty, pick default, then click Install Missing Custom Nodes again. See what to click.

Settings that worked, and what didn't

• Sampling: 8 steps, CFG 1, er_sde / simple, the Krea 2 Turbo recipe. The seed is fixed, so your first run matches the gallery.

• Canvas budget: keep it near 1.6 MP. At full size (budget 0) a 4 MP canvas painted its new sky a shade darker, creased at the corners and cut an invented hand at the old edge; 1 MP went blotchy. 1.6 to 2 MP stayed clean on every test photo.

• How far to grow: roughly doubling the picture works. Past that it starts to invent a second scene, so grow in two runs instead.

• Caption: keep Text Generate's thinking on. Krea 2's encoder has no thinking mode, and with it off the caption can come back empty.

• Inpaint in the same pass: AnyPaint treats any white mask pixel as "generate", so you can paint over something with the mask editor and pad at the same time.

• See what the LoRA carries: bypass it (Ctrl+B) and run again.

Speed

Measured on an RTX 5090 (32 GB) with the models already loaded: 14 s for the sample, 16:9 to 4:5 on the 1.6 MP canvas.

I haven't tested smaller cards yet. ComfyUI offloads what doesn't fit, so expect it to run slower rather than fail.

About the gallery

The PNG is a straight output of this workflow (seed 7701); drag it into ComfyUI to load its exact settings. The source photo was made with my Qwen Image 2.1 Studio workflow. The cover is cut from the real render.

License & credits

Krea 2 is by Krea and released under the Krea 2 Community License. Commercial use of the model, its derivatives or its outputs is allowed only if you (with your affiliates) make less than US$1M a year in revenue; above that, Krea asks for an enterprise license (krea.ai/krea-2-licensing). The Krea 2 files are Comfy-Org's ComfyUI repackage. The AnyPaint LoRA is by yijunwang2, under the same Krea 2 Community License. The workflow and the AusBoss nodes are free.

Changelog

• v1.0 (2026-09-25): first release.

Questions or bugs: GitHub issues or the comments here. Post what you make with it; I read everything. I post new workflows on X @Zanzibased and GitHub.