Getting to poster size without a second horizon

Two different problems

“I need a bigger image” hides two requirements that need different answers.

More pixels, same detail. A 1024 render that must print at A3. Nothing needs to be added; the existing picture needs to be larger. This is upscaling, and the finishing block covers what upscalers add and invent.

More detail at large size. A poster where the fabric texture, the foliage and the distant faces all need to hold up at 60 cm. This needs new detail generated, and that is what tiled diffusion is for.

Confusing them wastes time. If the image will be seen from two metres, it needs pixels, not detail.

Why you cannot just generate large

Covered in the brief block and worth restating: past its trained pixel count, a model composes badly: two horizons, duplicated subjects, a scene assembled from overlapping attempts. Memory is not the limit; composition is.

How tiled diffusion works

Ultimate SD Upscale, tiled diffusion and their equivalents all do the same thing:

1. Upscale the image conventionally to the target size: say 1024 to 4096 with an ESRGAN-family model.
2. Divide it into overlapping tiles, typically 512 or 768 with 64 pixels of overlap.
3. Run img2img on each tile at low denoise, 0.2 to 0.35, with a prompt.
4. Blend the overlaps.

The low denoise is what makes it safe. Each tile is being refined, not re-imagined, so no tile can decide there should be a second horizon in it. Push denoise above about 0.4 and tiles start inventing content, which produces the characteristic failure: a wall that becomes brickwork in one tile and plaster in the next.

Settings that matter

Denoise 0.2 to 0.35. The single most important value. Above 0.4 expect inconsistency.

Overlap 64 pixels minimum. Less and seams show. More is safer and slower.

A generic prompt. This is the setting people get wrong. Each tile gets the same prompt, and a tile containing only sky given a prompt about a woman in a sari will try to put her there. Use a short prompt describing the medium and finish (documentary photograph, fine grain, natural light) not the subject.

Seam fix if available, which runs an extra pass over the tile boundaries.

When not to use it

Faces. A face spanning two tiles gets two independent treatments and the result is subtly wrong. Repair faces at full resolution before tiling, and if a face still degrades, mask it out of the tiled pass.

Text, logos, product labels. Same problem, worse consequences. Anything that must remain exactly itself should be composited in after tiling, not put through it.

Large flat areas. Tiles introduce variation into flat surfaces, so a smooth studio backdrop acquires a faint patchwork. Mask it out or accept a plain upscale there.

The free path

All of this is free. Ultimate SD Upscale is an extension for A1111 and Forge; ComfyUI has tiled nodes in its core and its common node packs. The upscaler models themselves (ESRGAN, RealESRGAN, and the various community variants) are small files, run on modest hardware, and are free.

For the non-generative half, a plain upscale, GIMP’s Scale Image with LoHalo interpolation is decent, and the free ESRGAN command-line tools are better. Neither needs a GPU worth the name.

The limit, and the professional reality

Tiled diffusion adds plausible detail, not correct detail. At 4096 the fabric has weave, the wall has texture, the foliage has leaves; and none of it corresponds to anything. For a landscape or an atmospheric image that is exactly right. For a product where the client knows what the label says, it is a fabrication that gets more detailed the larger you print it.

Which produces a rule worth holding: the parts of the image that must be correct do not go through a generative upscale. Composite them in at full resolution afterwards.

The other honest point is that large-format work rarely needs this at all. Viewing distance does most of the job. A roll-up banner read from two metres is fine at 150 dpi; a billboard at 30 metres is fine at 15. Before spending forty minutes on a tiled pass, calculate what the eye can actually resolve at the viewing distance: it is frequently less than the file you already have, and the whole exercise turns out to be unnecessary.