Detail SEGS by Scale Factor w/ Tiled Diffusion

Here's the gap in the plain scale-factor detailer: it resamples each crop at full size, which is exactly what you want for a small face - and exactly the wrong thing when a detected region is huge. Detail a region that occupies half a 4K frame and the crop alone can exceed your VRAM. Detail SEGS by Scale Factor w/ Tiled Diffusion exists for that case: it's the same crop-refine-paste loop, but the resampling step uses tiled diffusion, so each crop is evaluated in overlapping tiles instead of as one giant latent.

How it works

Everything from Detail SEGS by Scale Factor is here - image, segs, model, vae, a conditioning batch matched one-to-one against SEGS order, scale_factor (1.5), upscale_method (lanczos), clamp_size, sampler controls, feather, noise_mask, tiled_encode/tiled_decode. What's added is the tiled sampling stack:

• diffusion_mode: multidiffusion (default, averages overlapping predictions - steady) or Mixture of Diffusers (weights tile centers more strongly - softer blend).
• latent_tile_width / latent_tile_height (128×128): tile size in latent pixels. Bigger tiles see more context, cost more memory.
• latent_tile_overlap (16): overlap between tiles; more overlap kills seams, costs compute.
• latent_tile_batch_size (4): tiles sampled per batch; higher is faster, thirstier.

Output is one image with the detailed regions blended back.

When to reach for it

Use this over the plain version when the region is large enough that a full-res crop would be a problem - big background objects, wide full-body regions, anything where the crop approaches or exceeds the sampler's comfortable resolution. The tradeoff is honest: tiled sampling fixes the memory problem but not the comprehension problem. Tiles each see a piece of the crop, and on very large canvases the pieces can lose track of the whole - that's precisely the failure mode SimpleSyrup's Contextual Diffusion was invented to fix, and the README notes this node shares the tiled sampling path with the tiled samplers rather than adding whole-image correction. If you're detailing a crop where composition within the crop matters, keep the tile size generous and the overlap up.

Also worth remembering for any detail pass: do it before upscaling, don't stack passes, and watch for the grey-box seam if the composite edges look off (raise feather).

Install

Ships in the SimpleSyrup pack. ComfyUI Manager → search SimpleSyrup → Install → restart, or:

Set-Location ComfyUI\custom_nodes
git clone https://github.com/Artificial-Sweetener/SimpleSyrup.git
Set-Location SimpleSyrup
..\..\venv\Scripts\python.exe -m pip install -r requirements.txt

Restart ComfyUI; needs a current build (v3 extension API).

Common issues

• Still out of VRAM - lower tile size or latent_tile_batch_size; those are the two VRAM levers.
• Seams inside the detailed crop - raise latent_tile_overlap, or switch to Mixture of Diffusers.
• Region looks locally coherent but composition drifts - that's the tiling comprehension limit; bigger tiles help, and for scene-critical regions consider Detail SEGS as Regions or Contextual Diffusion.
• Node missing - update ComfyUI first.

It's the detailer for the big-canvas end of the scale-factor family: same loop, tiled guts, one more option when the crop is the problem.