A LoRA for FLUX 3 Image Klein that extends a picture on any side, and paints the gray corners of a picture you have turned. Pad the picture with flat gray, run it, and the gray is filled with more of the scene while your picture stays where it was.

The workflow for it: FLUX 3 Image Klein Outpaint. It pads or turns your picture, writes the prompt and puts your original back on top. Also for FLUX 3 Klein: Single & Multi-Image Edit.

FLUX 3 Image Klein often zooms in on the picture or slides it instead of extending it. In my test of the official template graph it did that in 36 of 40 runs. With this LoRA, the original part of the picture moved by 0.1 to 1.5 px on six test pictures it was not trained on. On 16 more test pictures it stayed within 2 px on 15 at strength 1.5, and on 12 at 1.0.

Use it on the distilled model, at strength 1.5, with gray #808080 and the instruction below. I trained it on the base model and tested it on the distilled one. Other fills and prompts are untested.

How to use it

The workflow file in the Hugging Face repo does all of this: load your image and press Run. To build it yourself:

1. Pad your picture with flat gray #808080, each side a multiple of 32, about 1 MP. My Image Crop + Rotate + Pad node does it, with feather 0. It also turns a picture and leaves gray corners.

2. Prompt: the instruction first, word for word. After it you can add Scene: and a description of the picture.

3. Set Reference Multi Latent with the padded picture as the reference, resolution keep resolution. Sample on the latent it returns: 16 steps, CFG 1, euler / simple, denoise 1, the LoRA at 1.5.

4. VAE Decode, then Stitch Inpaint from my nodes with the seam classic and Tone match 1. It puts your original pixels back on top.

The instruction is the trigger:

Outpaint the image: replace the solid gray areas with a seamless continuation of the scene, keeping the existing picture unchanged.

Don't add a noise mask. Plain FLUX 3 needs one to hold the picture in place, but then the brightness of your picture shifts and a box can show. With the LoRA you don't need it.

What you need

• ComfyUI 0.41.0 or newer. That is the first version with FLUX 3.

• AusBoss nodes 2.13.0 or newer, for Image Crop + Rotate + Pad and Stitch Inpaint. In ComfyUI-Manager, search "AusBoss". Already have them? Update first.

• The FLUX 3 Image Klein files. They are on Hugging Face behind a license click: open black-forest-labs/flux-3-image-klein-comfy, sign in and accept, then these links work.

  • flux3_image_klein_sd_int8_convrot.safetensors → models/diffusion_models · 5.83 GB

  • qwen3vl_4b_int8_convrot.safetensors → models/text_encoders · 4.86 GB

  • flux2-vae.safetensors → models/vae · 336 MB

• This LoRA → models/loras · 161 MB

The LoRA works the same on the INT8 files and the full-size BF16 files, and INT8 took about half the time in my check.

Tips

• Keep the fill and the feather. Gray #808080 and feather 0 are what it learned: a hard gray edge. Another colour is untested.

• Pull out about your picture's own size at most. It was trained with your picture covering 45 to 93 % of the canvas. For more, run it again on the result.

• Paste the original back. FLUX 3 shifts the colours of the whole picture a little, mostly towards red. Stitch Inpaint puts your pixels back, and Tone match makes the new area follow your picture's tone.

• Not happy with the new area? Run it again with another seed. In my trial runs, one in 18 drew part of the picture a second time, and at 1.5 one in 20.

• About 8 seconds a run on an RTX 5090 with the INT8 files.

Training

This video explains how I trained it: youtube.com/watch?v=5YN7ywxS4I8. It's about the Qwen Image 2.1 version, but I used the same technique for this one, and the workflow in the video is basically the same one too.

• ai-toolkit (flux3) on FLUX 3 Image Klein Base in BF16. Rank 32, alpha 32, AdamW 8-bit, learning rate 1e-4, batch 1. This is step 1,000 of a 2,000-step run. I compared steps 750, 1,000 and 1,500 on six pictures it never saw, and liked 1,000 best.

• 102 pairs: 68 outpaint pairs from the pictures and captions of my Qwen Image 2.1 Outpaint LoRA v2, and 34 turned pairs from the same pictures. The target is the real photo, the source is the same photo with the area to paint filled with #808080.

• Pictures: Unsplash photos under the Unsplash License, museum pictures from PD12M, anime from anime-with-caption-cc0, and some from my own collection.

The test numbers, how each step compared and the rest of the notes are on the Hugging Face model card.

On Qwen Image 2.1 instead: Qwen Image 2.1 Outpaint LoRA.

License

FLUX 3 Image Klein is by Black Forest Labs, under the FLUX Kommunity License v.1.0. This LoRA is a modified version of it: an adapter I trained on FLUX 3 Image Klein Base. I share it under the same license and added no terms of my own. The rights to use FLUX 3 Image Klein and this adapter come to you directly from Black Forest Labs under that license, and a copy of it is in the Hugging Face repo. Whether you may use the pictures you make commercially depends on section 2 of it, so read it first.

Black Forest Labs did not make, endorse or approve this adapter. It is provided as is, with no warranty and no liability, as in sections 5 and 6 of the license.

The notice the license asks for:

This FLUX Model is licensed by Black Forest Labs Inc. under the FLUX Kommunity License. Copyright Black Forest Labs Inc. IN NO EVENT SHALL BLACK FOREST LABS INC. BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.

I changed the FLUX Model: this file is a LoRA adapter trained by ausboss.

v1.0 (2026-10-15): first release (step 1,000).

Questions or bugs: the comments here or GitHub issues. Videos on YouTube @ausbossdev. More from me on X @Zanzibased and GitHub.