Extend Any Photo to a New Shape With FLUX 3 Image Klein and a Flat Grey Border
Ask an AI image editor to make a square photo taller and it will usually do something you did not ask for. It zooms in. It nudges the subject. It redraws the face slightly, because it has decided the whole picture is up for negotiation. You wanted more sky above a portrait and you got a different portrait.
The fix that turned up a few hours after Black Forest Labs released FLUX 3 Image Klein on October 15 is almost comically low-tech. Paint the area you want filled a flat, boring grey. Hex code #808080, nothing else. Then load a small free add-on, trained by a developer who goes by ausboss, that has learned one rule: grey means "make something up here," and everything that is not grey means "leave it alone."
That single rule is the whole trick, and it explains both why the method works and where it fails. If you produce content in more than one shape, a 4:5 feed post, a 9:16 story, a 16:9 banner from the same shoot, this is worth an afternoon.
Why extending a picture is harder than editing one
Most AI editing assumes the picture is a suggestion. You describe what you want and the model repaints, trying to stay close to the original. For changing a jacket colour, that is fine. For extending a frame, it is the wrong mindset, because the original is the one part that must not change. The client approved those pixels.
The add-on's author reports that without it, Klein often zooms in on or shifts the picture when asked to extend it. That matches what anyone who has tried this with other models has seen. A model that was never told which part is sacred will treat all of it as editable.
The grey border solves the communication problem. A flat field of one exact colour almost never appears in a real photograph, so it works like a highlighter. The add-on was trained on pictures padded this way, so it learned to read grey as "the empty part of the canvas" and to continue the scene into it.
Then the workflow adds a safety net. After the model paints, a node called Stitch Inpaint puts your original pixels back on top, so even if the model drifted slightly inside your picture, the final file carries your real photo. It also matches the tone of the new area to your picture, so the join does not show as a change in brightness.
That is a sensible design. It does not trust the model with the part that matters.
The number that decides whether it works
The most useful line on the add-on's page is not the prompt. It is this: the add-on "was trained with your picture covering 45 to 93 % of the canvas."
Translate that into the jobs you actually do:
- A 4:5 portrait post (1080 by 1350) extended to a 9:16 story (1080 by 1920). Your photo covers about 70 percent of the new canvas. Comfortably inside the range.
- A square photo extended to 9:16. Your photo covers about 56 percent. Still inside.
- A 4:5 portrait extended sideways into a 16:9 banner at the same height. Your photo covers about 45 percent. Right at the edge, and you should expect more retries.
- A tight crop extended to a wide establishing shot. Under 45 percent. Outside what it learned.
For that last case, the author's advice is to go in steps: "For a bigger extension, run it again on the result." Extend part of the way, check it, then extend the result. Each pass gives the model a bigger, more complete scene to continue from.
This one number is more useful than any prompt advice, because it tells you before you start whether you are asking for something the add-on was built to do.
Put this into practice
You need a computer that can run FLUX 3 Image Klein in ComfyUI, the free node-based app most local image work runs in. Black Forest Labs says compressed versions of Klein run on NVIDIA cards with as little as 8 GB of memory. If you do not have that, you can still follow the method in principle, but this add-on is a ComfyUI tool.
1. Get the pieces. Update ComfyUI to version 0.41 or later. Open ComfyUI-Manager, search for "AusBoss," and install the AusBoss node pack, version 2.13.0 or later. Download the add-on and its ready-made workflow from the Outpaint add-on page on Hugging Face. The workflow file is workflows/flux-3-image-klein-outpaint.json; drag it onto the ComfyUI canvas to load it. You also need Klein's own files, which BFL publishes in a ComfyUI repo after you accept the licence.
2. Decide the new shape and check the coverage. Work out how much of the new canvas your original will cover. If it is between roughly half and nine tenths, do it in one pass. If it is less, plan two passes.
3. Let the workflow pad the picture. The included workflow has a padding node already set to fill with #808080 and no soft edge. Set where the new space goes: above, below, to the sides. The workflow sizes the working canvas to about one megapixel (roughly 1000 by 1000 pixels, or the same area in another shape), which is the size the author tested.
4. Write the prompt in the author's order. Start with the instruction, word for word: "Outpaint the image: replace the solid gray areas with a seamless continuation of the scene, keeping the existing picture unchanged." Then, if you like, add a line beginning Scene: that describes what should be in the new space. "Scene: open sky with high thin clouds, late afternoon light" does more than you would expect, because the model no longer has to guess.
5. Keep the author's settings. 16 steps, the fast version of Klein, add-on strength 1.5. The author says 1.0 also works but fails more often when the grey area is large.
6. Run four, keep one. Generate a few versions with different random seeds (the number that makes each attempt come out differently). The author's own failures were all of the kind that a second try fixes, and running a few at once is cheaper than fixing one by hand.
7. Check the seam at full size. Zoom to 100 percent on the line where your photo ends and the new area begins. Look for repeated objects, a change in grain, and lines (horizons, table edges, railings) that do not continue straight across.
8. Finish in your usual editor. The add-on, in the author's words, "does not make the picture bigger or sharper." If you are delivering a large file, the new area may need upscaling or a grain pass to match the original.
Where it breaks
The author's limitations section is unusually specific, which is a good sign. Here it is in plain terms.
It repeats things. In the author's 18 trial runs, one drew part of the picture a second time. That happened with a large grey area above a busy, rotated picture. In another test at the recommended strength, a seated person was drawn slightly lower in the new area and, after the original was pasted back, "the paste-back showed her knees twice." This is the classic outpainting failure, and it is most likely when the subject touches the edge you are extending from.
Seams can show as a light patch. One run left a lighter strip of ceiling at the join. The tone-matching helps, but it is not perfect.
It is tested at about one megapixel. The add-on was trained on canvases from roughly one to two megapixels, and the author tested about one. Anything larger is untested. For print or large display work, extend at this size and upscale afterward.
It was only tested on the fast version of Klein. If you use the slower Base model (the version meant for training add-ons), you are in territory the author did not check.
Subjects at the edge are risky. If a person's arm or a product's corner is cut off by the frame, the add-on has to invent the rest of it. That is the hardest thing to get right, and the place where hands and labels go strange. Prefer extensions on the side away from the subject.
The licence follows the model. The add-on uses the same FLUX Kommunity License as Klein. Free commercial use for individuals and companies under US$5 million in annual revenue; above that, you need a licence from Black Forest Labs. The author notes the add-on is not endorsed by BFL.
It is one person's add-on, one day old. There is no company behind it and no support desk. It may change, and the node pack may change with it.
Where this leaves you
Every format change used to mean either a reshoot, a careful crop that cost you composition, or an hour with a clone tool. AI outpainting promised to end that and mostly delivered a different picture. This add-on gets closer to the real job because it changes the question. It stops asking the model to reinterpret your image and asks it only to continue it, and then puts your pixels back to be sure.
The habit worth taking from it applies beyond Klein: before you extend anything, work out how much of the new frame your original covers. Above half, go ahead. Below half, go in steps. Try it on one shoot you already delivered in a single shape, and see how many of the other shapes you could have sent the same day.
Medium metadata
Title: Extend Any Photo to a New Shape With FLUX 3 Image Klein and a Flat Grey Border
Subtitle: A free add-on for Black Forest Labs' new open image model fills grey padding and leaves your original untouched, so one shoot can become a story, a feed post and a banner. Here is the workflow and the one number that decides whether it works.
Tags: Artificial Intelligence, Photography, Social Media Marketing, AI Art, Content Creation
Estimated read time: 8 minutes