Extend a picture on any side, and FLUX.2 Klein 9B paints the new area with Pixaroma's PixaOutpaint LoRA. Your image goes back on top. Only the new area and a thin strip along the old edge change.
How to use it
Load your image.
Pick a ratio, or drag the orange handles out. The small button next to the ratios flips portrait and landscape.
Optional: describe the new area in What goes there, after the first sentence.
Press Run.
What you need
AusBoss nodes 2.5.0 or newer. Already have them? Update first.
The Klein 9B models, the PixaOutpaint LoRA and the caption model:
flux-2-klein-9b-int8-convrot.safetensors →
models/diffusion_models· 9.43 GBqwen_3_8b_fp8mixed.safetensors →
models/text_encoders· 8.66 GBflux2-vae.safetensors →
models/vae· 336 MBpixaoutpaint.safetensors →
models/loras· 127 MBqwen3vl_8b_int8_convrot.safetensors →
models/text_encoders· 9.35 GB (it only writes the description)
The sample picture is in
sample_inputs.zip. Put it inComfyUI/input, or load your own image.
Tips
Keep the first sentence in What goes there. It holds the LoRA's trigger. Without it the new area can come out blank.
Crop out edges that confuse it. Noise, a frame line or a busy strip along the edge of your picture can make the new area come out strange. Drag the blue handles on Image Crop + Rotate + Pad inward to cut those edges off first.
Tight portraits: say what's below after the first sentence, like "the rest of her body".
A big pad can grow a second hand or face near your picture's edge. If it does, extend in two runs so your picture fills at least half the canvas each time.
The result is about 1.6 MP, whatever size your picture is. A 5000 px photo comes back about 1120 × 1488.
Leave Align at 1. A bigger Align adds pixels on the right and bottom, even on a side you didn't pad.
About 8 seconds a picture on an RTX 5090 once the models are loaded, 4 seconds for a new seed.
Credits
FLUX.2 [klein] 9B is by Black Forest Labs, under the FLUX Non-Commercial License: non-commercial use only. The INT8 file is a community conversion by Winnougan; the text encoder and VAE are Comfy-Org's repackage. PixaOutpaint is Pixaroma's experimental outpaint LoRA for Klein 9B (Pixaroma/experimental_loras, MIT). The caption model is Qwen Image 2.1's Qwen3-VL 8B text encoder, repackaged by Comfy-Org, under the Qwen Research License (non-commercial use only).
v1.1 (2026-09-30): the padding follows your picture, the seam blends in, and tight portraits stopped getting a second face in my tests. v1.0 (2026-09-25): first release.
Questions or bugs: GitHub issues or the comments here. More from me on X @Zanzibased and GitHub.
Description
The padding now follows your picture. Pick a ratio or drag the orange handles, and it's worked out for whatever image you load. v1.0 saved it in pixels for the sample, so other pictures could get far too much.
Fixed the second face under tight portraits. The description used to describe the person, and Klein painted them again under the join. Now it describes only the setting.
Image Crop + Rotate + Pad replaces Load Image + Pad, which gives you the ratio buttons.
The workflow is smaller: 14 nodes instead of 24, with the prompt and the painting in two boxes.
The seam is set to blend in, with Tone match on. The old seam could turn the new area's blacks a hazy gray-green, or leave a line on a turned picture.
Needs AusBoss nodes 2.5.0 or newer. The models are the same.
FAQ
Comments (2)
Tested v1.1 of the workflow, nice addition of the Crop/Rotate/Pad node here :) , here are few notes:
As I don't have Qwen image 2.1 & I already have several Qwen VLMs between 4B & 3.5 9B (all are uncensored anyways) so I just used them to read the image & feed their response to "Your line, then the description" node, the outpaint results are very good with high success rate.
I recommend those Krea2 & Klein workflows & your LoRAs for people who wanna save time & effort instead of longer manual engineered prompting each single time or instead of crowded complex workflows. But,, people need to either use a QwenVL node or any method to make a vision model reads the image "in case they don't own the encoder of Qwen Image 2.1", (relying only on the main simple prompts won't give the desired final results at all)
good notes, thanks. the description line in that workflow is written by the qwen3-vl 8b node, not the plain prompt boxes. you're right the page doesn't say that clearly though.