NSFW AI Outpainting Guide: Extend Any Scene (2026)

Outpainting extends an AI image past its original borders, letting you change the aspect ratio, zoom out, or reveal more of a body or scene that was cropped. You pad the canvas with empty pixels, mask that new area, and let the model generate into it at moderate denoise (around 0.7 to 0.85) with a few pixels of mask blur to hide the seam. A1111, Forge, and ComfyUI all support it.

What outpainting actually does

Inpainting fills a hole inside an image. Outpainting does the opposite: it generates new content outside the current frame. You start with a finished render that is cropped too tightly (a portrait that cut off the legs, a square that you now want in 16:9, a close-up you wish was a wider shot) and the model paints plausible new pixels into the blank margin you add around it.

The mechanism is the same diffusion process used everywhere else. You enlarge the canvas, the new region is filled with neutral or noise pixels, a mask marks that region as “regenerate,” and the sampler denoises it while looking at the original content for context. The trick that makes outpainting hard is continuity: the new pixels must match the lighting, skin tone, perspective, and style of the original, with no visible line where old meets new.

For NSFW work specifically, outpainting solves two recurring annoyances. First, models frequently crop bodies awkwardly, cutting off feet, hands, or the lower body. Second, the aspect ratio you generated at is rarely the aspect ratio you need for a wallpaper, a banner, or a phone screen. Outpainting fixes both without regenerating from scratch and losing the face or pose you already liked.

If you have not built a local Stable Diffusion setup yet, you can prototype compositions with our free NSFW AI image generator in the browser first, then bring a result you like into a local install to extend it. Outpainting itself needs a local tool, since it depends on masked diffusion controls that browser widgets do not expose.

A1111: outpainting mk2 and poor man’s outpainting

Automatic1111 ships two outpainting scripts in the img2img tab, found under the Script dropdown at the bottom: “Outpainting mk2” and “Poor man’s outpainting.”

Send your image to img2img, then choose a script:

Script: Outpainting mk2
Pixels to expand: 128  (per selected direction)
Mask blur: 8
Directions: up / down / left / right (check the sides to grow)
Denoising strength: 0.8
Sampling steps: 30 or more

Outpainting mk2 uses a noise-and-fill approach with a falloff and color-variation control that generally produces smoother extensions than the simpler poor man’s script. Poor man’s outpainting is more literal: it pads, masks, and regenerates with your normal masked-content settings. Mk2 is the better default for skin and gradients; poor man’s can be steadier for hard geometric backgrounds.

Expand in modest increments. Asking for 512 pixels of brand-new body in one pass invites anatomy errors and repeated elements. Add 128 to 256 pixels per pass, accept the result, then send the extended image back to img2img and expand again. Iterative outpainting beats one giant jump almost every time.

Keep your prompt aligned with what should appear in the new area. If you are extending downward to add legs, your prompt should describe the full standing figure, not just the face that was already there. The model paints what the prompt asks for in the masked region, so a face-only prompt extended downward tends to grow a second head or a torso that makes no anatomical sense.

Forge outpainting

Forge keeps the same img2img scripts as A1111 and behaves almost identically, so the mk2 settings above carry over directly. The practical differences are speed and memory: Forge’s optimized backend handles the larger canvases that repeated outpainting creates with less VRAM pressure, which matters once you are working at SDXL resolutions and pushing a 1024 image out to 1536 or beyond.

If you run a low-VRAM card, Forge is usually the more comfortable home for outpainting because each pass grows the working resolution. A 6GB or 8GB GPU that handles a single 1024 render fine can stall when that canvas balloons across three outpaint passes. Our roundup of low-VRAM NSFW checkpoints pairs well with Forge for this exact reason. Tile-based VAE decoding, which Forge enables readily, also helps keep memory flat as the canvas grows.

ComfyUI: the pad-and-inpaint approach

ComfyUI does not have a one-click outpaint button; you build it from nodes, which is more work but far more controllable. The canonical pattern uses the Pad Image for Outpainting node feeding an inpainting workflow:

Load Image
  -> Pad Image for Outpainting (left/top/right/bottom in px, feathering)
  -> VAE Encode (for Inpainting)  [takes image + mask]
  -> KSampler (denoise 0.8, your checkpoint + prompt)
  -> VAE Decode -> Save Image

The Pad Image for Outpainting node adds the blank margin and outputs a matching mask automatically, with a feathering parameter that softens the boundary (the ComfyUI equivalent of mask blur). Feed both into a VAE Encode for Inpainting node so the sampler knows which region is fixed and which is free.

The big advantage in ComfyUI is the inpaint model option. You can run a dedicated inpainting checkpoint, or use the Fooocus-style inpaint patch nodes, which were trained to extend content and tend to handle large outpaints with fewer seams and fewer hallucinated objects. For complex multi-pass extensions, you can chain two pad-and-sample groups in one graph and run the whole thing as a single click. Our full ComfyUI for NSFW guide covers building and saving these graphs.

Settings reference for clean extensions

The values below are reliable starting points. Adjust denoise down if the new area drifts away from the original style, and up if it comes out blurry or too faithful to the padded gray.

Setting	Recommended range	What it controls
Pixels to expand per pass	128 to 256	Smaller is safer; iterate rather than jump
Denoising strength	0.7 to 0.85	How freely the model fills the new region
Mask blur / feathering	8 to 24 px	Hides the seam; too low leaves a hard line
Sampling steps	30 to 40	Detail in the generated margin
CFG scale	5 to 7	Lower keeps continuity, higher follows prompt harder
Passes	2 to 4	Build the full extension gradually

These are starting points, not laws. SDXL-based models (including Pony and Illustrious) usually want the denoise toward the lower end of the range because their stronger priors fill the gap more confidently; SD1.5 often needs a touch more denoise and more mask blur to blend skin smoothly.

Keeping continuity and fixing seams

Continuity is the whole game. A few habits keep the extension believable:

Match the prompt to the full intended frame. Describe the complete figure and scene you want to end up with, not just what was in the crop. The model only has your prompt plus the existing pixels to reason from.

Use enough mask blur. A hard mask edge with no blur leaves a visible line. Twelve to twenty pixels of blur usually dissolves it. If the seam still shows, the cure is often a small follow-up inpaint pass straddling the boundary at low denoise (around 0.3 to 0.4) to harmonize the two sides.

Lower CFG for smoother joins. High CFG makes the model assert the prompt aggressively, which can produce a sudden change in tone at the seam. Dropping CFG to 5 or 6 lets it defer more to the existing content.

Watch for repetition. Outpainting large empty areas (a plain wall, bedsheets, sky) sometimes tiles the same texture. Break it with a descriptive prompt for the background, or extend in smaller steps so each pass has more real context to anchor to.

Do a continuity inpaint after. Even a clean outpaint benefits from a final low-denoise inpaint over the transition zone to fix any subtle tone or lighting mismatch. See our inpainting guide for the masking workflow, and our img2img guide for the broader refine-and-extend mindset.

Common outpainting mistakes

Expanding too far in one pass. The single biggest cause of garbage outpaints. Half the time the model invents a duplicate torso or an extra limb because you asked it to imagine too much at once. Cut the pixel count and add a pass.

Forgetting to update the prompt. Extending downward with a head-and-shoulders prompt grows nonsense. The prompt must describe what belongs in the new region.

Mask blur too low. A crisp seam line is almost always a blur problem. Raise it.

Wrong denoise. Too low and the new area stays gray and undefined; too high and it ignores the original style entirely. Stay in the 0.7 to 0.85 band and adjust in small steps.

Anatomy at the new edge. When you reveal hands or feet that were originally cropped, expect to clean them up. Hands especially almost always need a dedicated pass; our fix hands guide covers the ADetailer and inpaint approach.

Outpainting versus generating wide from the start

A fair question is why bother outpainting at all, rather than just generating in the wide aspect ratio you want from the beginning. The answer is composition control and preservation. When you generate a fresh image at 16:9, you get whatever the model decides to put in that wide frame, and the figure may be small, off-center, or posed differently than you wanted. Outpainting lets you nail the subject first at a comfortable ratio, then build the rest of the scene around a result you already approved.

There is also the matter of saving good renders. You will frequently get a face, expression, or pose that is exactly right but framed too tight. Regenerating risks losing that lucky combination forever. Outpainting keeps the part you love untouched and only invents the margins. For NSFW work, where a specific pose or body type can take many seeds to land, that preservation is genuinely valuable. Think of outpainting as a way to protect your best output while still reshaping the frame.

The flip side: outpainting is slower and more iterative than a single generation, and it can introduce continuity work that a clean wide render would not. The rule of thumb is simple. If you do not yet have a subject you care about, generate wide and pick a good one. If you already have a keeper that is just framed wrong, outpaint it. Both approaches live in the same finishing pipeline and complement each other.

Model choice and outpainting quality

The checkpoint you outpaint with matters more than people expect. A model with strong, coherent priors fills empty margins more confidently and hallucinates fewer stray objects. Photoreal SDXL checkpoints tend to extend skin, fabric, and environments cleanly, while some heavily stylized anime checkpoints can struggle to continue a complex background and will smear or repeat textures. If your outpaints keep producing mush in large empty areas, trying a different checkpoint for the extension pass is a legitimate fix; you can outpaint with one model and do your detail passes with another.

Dedicated inpainting checkpoints, or the inpaint-patch approach in ComfyUI, almost always beat a standard checkpoint for large extensions because they were trained specifically to continue existing content rather than generate from scratch. If outpainting is a regular part of your workflow, keeping one good inpainting model installed is worth the disk space. Our guides to the best NSFW checkpoints and how to install checkpoints cover sourcing and setup. Matching the outpaint checkpoint to the original render’s style also keeps the seam invisible, since both halves share the same texture and color characteristics.

A practical outpainting routine

Here is a routine that consistently produces clean extensions:

1. Pick the final aspect ratio you want and figure out which directions need to grow.
2. Expand one or two directions at a time, 128 to 256 pixels each, never all four sides at full reach in one shot.
3. Set denoise 0.8, mask blur 12 to 16, CFG 5 to 6, 30+ steps.
4. Rewrite the prompt to describe the full intended frame, including the new content.
5. Generate, accept the best of a small batch, then send the result back in and repeat until you hit the target dimensions.
6. Run a low-denoise continuity inpaint across each seam.
7. Finish with face and hand cleanup, then upscale.

That last sequence (extend, harmonize, fix details, upscale) is the same finishing chain every advanced image goes through. You can prototype the look in the browser generator before committing to a local outpaint session, but the extension itself rewards patience and small steps. Treat each pass as a small, verifiable improvement rather than a single magic expansion, and your success rate climbs dramatically. The images that look effortlessly wide or full-body almost always got there through three or four modest passes, not one heroic jump.

When outpainting is the wrong tool

Outpainting is not a cure for everything. If the original render has a broken pose, mangled anatomy in the existing frame, or a face you do not actually like, no amount of extending the borders will fix it. Outpainting only adds to what you have; it does not repair what is already there. For internal defects, inpainting and a regeneration are the right tools, not outpainting.

It is also the wrong choice when you need a fundamentally different composition. If you want the subject standing instead of sitting, or facing the camera instead of away, outpainting cannot deliver that; it would have to invent an entirely new pose in the margins, which it does poorly. In those cases, change the prompt or the ControlNet pose and regenerate. Reserve outpainting for the specific job it is good at: extending a frame you are otherwise happy with, in directions where the new content is a natural continuation of what already exists. Used that way, it is one of the most reliable finishing tools in the kit. Used as a band-aid for a fundamentally weak render, it disappoints every time.

Frequently asked questions

What denoising strength should I use for outpainting?

Stay in the 0.7 to 0.85 range. Too low and the new region stays gray and undefined because the model barely changes the padded pixels; too high and it ignores the original style, producing a mismatched extension. SDXL-based checkpoints like Pony and Illustrious usually want the lower end around 0.7, while SD1.5 often needs slightly more denoise plus extra mask blur to blend skin smoothly.

Why does my outpainted area have a visible seam?

A hard line where old meets new is almost always insufficient mask blur or feathering. Raise it to 12 to 24 pixels. If a faint mismatch in tone or lighting remains, run a follow-up inpaint pass straddling the boundary at low denoise around 0.3 to 0.4 to harmonize both sides. Lowering CFG to 5 or 6 during the outpaint also reduces sudden tonal jumps at the join.

How far can I extend an image in one pass?

Keep each pass to 128 to 256 pixels per direction. Expanding 512 pixels of brand-new content at once invites anatomy errors, duplicated torsos, and tiled textures because the model has to invent too much with too little context. Iterative outpainting, where you extend, accept, then send the result back and extend again, is far more reliable than one large jump.

Does ComfyUI have a one-click outpaint button?

No. ComfyUI builds outpainting from nodes, typically Load Image into Pad Image for Outpainting, then VAE Encode for Inpainting, KSampler at around 0.8 denoise, and VAE Decode. The Pad node adds the margin and outputs a matching mask with feathering. This is more setup than A1111 scripts but more controllable, and it lets you use dedicated inpaint models or Fooocus-style inpaint patches for cleaner large extensions.

Why does my prompt matter when outpainting downward?

The model paints whatever the prompt describes into the masked region, using the existing pixels only as context. If you extend downward to add legs but your prompt only describes a face, the model may grow a second head or an incoherent torso. Always rewrite the prompt to describe the full intended frame, including the body parts or scenery that should appear in the new area.

What is the difference between Outpainting mk2 and Poor man’s outpainting in A1111?

Both are img2img scripts. Outpainting mk2 uses a noise-and-fill approach with falloff and color-variation controls, producing smoother extensions for skin and gradients, and is the better general default. Poor man’s outpainting is more literal: it pads, masks, and regenerates using your standard masked-content settings, which can be steadier for hard geometric backgrounds but rougher on organic surfaces.

Is Forge better than A1111 for outpainting?

Forge shares the same img2img outpainting scripts, so results are nearly identical, but its optimized backend handles the growing canvas with less VRAM pressure. Since each outpaint pass enlarges the working resolution, Forge is more comfortable on low-VRAM cards and at SDXL resolutions. Enabling tiled VAE decoding in Forge keeps memory flat as the canvas expands across multiple passes.

How do I change aspect ratio with outpainting?

Decide the target ratio first, then identify which sides need to grow. To turn a square into 16:9, extend left and right; to turn a portrait into a fuller-body shot, extend downward. Expand those directions in 128 to 256 pixel increments over several passes, updating the prompt to describe the widened scene, and finish with a continuity inpaint across each seam before upscaling.