KSampler (Contextual Diffusion)

Tiled diffusion has a dirty secret. It keeps each evaluation small enough to fit in VRAM, but every tile sees only a piece of the image - so on a big canvas, tiles quietly invent their own version of the scene. The pack author hit this hard editing a 2160×3072 photo with FLUX.2 Klein: one tile found a figure, another invented a second figure, and different parts of a cathedral became different buildings. The overlaps were seamless and the scene was nonsense. KSampler (Contextual Diffusion) is the original method they built to fix that, and it's the flagship node of SimpleSyrup.

How it works

The idea is deceptively simple: while tiled local predictions handle full-resolution detail, a second, smaller, aspect-preserving view of the whole image keeps the composition honest. The whole-image prediction is upsampled and blended into the tiles - but not by direct averaging, which produced blur, and not for the whole run, which smeared detail. What worked is subtracting the low-frequency interpretation already present in the tiled prediction and adding only the difference:

prediction = local + scheduled_weight × (global_upsampled − local_low_frequency)

In plain terms: the global branch corrects composition early, then hands control back to the tiles for fine texture before it can smear anything. The author acknowledges this is closely related to Upsample Guidance (arXiv:2404.01709), developed independently - a multiscale low-frequency-guidance idea.

The controls map onto that story:

• global_weight (1): how authoritative the whole-image guidance is. Lower it to let tiles interpret more freely.
• global_steps (1): how many initial steps use the global context. For distilled models that finish in four steps, even one corrected step is a quarter of the process - hence the defaults of steps 4, cfg 1.
• global_decay (0.5): how fast the whole-image strength fades after each global step. Lower = tiles take over sooner.

Plus the tiled plumbing: diffusion_mode (multidiffusion or Mixture of Diffusers), latent_context_size (96), latent_context_overlap (32), latent_context_batch_size (4). There's an optional segs input that replaces the regular grid with SEGS-guided context windows - a way to steer where the local evaluations land - and a contexts_segs output that shows you the windows that actually ran (masks are only built when you connect it, so don't leave it dangling if you're tight on VRAM).

What it's for - and what it isn't

Contextual Diffusion is for edits the model already knows how to make at normal resolution: clothing, color, material, jewelry, expression, local lighting - changes where you want to keep the source pose and composition. It is not for inventing a new pose, camera, or environment; establish those at a normal resolution first, then refine. Honest limitations: UniPC, regional conditioning, ControlNet, and GLIGEN are rejected because the author hasn't validated their spatial behavior across both context sizes. If you need per-region prompts on a big canvas, that's KSampler (Prompt by Tiled Region); if you need plain tiling with no whole-image correction, that's KSampler (Tiled Diffusion).

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; current build required (v3 extension API).

Common issues

• Blurry result - the whole-image branch overstayed. Lower global_steps or global_decay.
• Smears / repeated edges - global correction active too long at high weight. Reduce global_steps or global_weight.
• Composition still drifts - raise global_weight and global_steps; you went too far the other way.
• Nodes missing - update ComfyUI first.

It's the most interesting node in the pack: a genuinely new sampler method, documented honestly, built for the exact case where ordinary tiled diffusion fails.