Near-sighted Attention (Simple)

This is the variant for people who looked at the full Near-sighted Attention node, counted the inputs, and closed the tab. Simple does the same tiling-and-pooling combo - split self-attention into tiles, mix in a coarse global view, thin the keys and values - but the fiddly plumbing is hard-wired. Modes are fixed to 2D average pooling, depth decay is fixed, and the separate per-key/per-value blend knobs are gone. What's left is one slothful rate and two blend sliders. Same mechanism, same MODEL in → MODEL out, a fraction of the staring-at-sliders time.

What you set

From the node's schema, the exposed controls are mostly the tiling side plus three style values:

• base_tile_size / peak_tile_size (24 / 32) - tile size in latent space, interpolated across the sampling schedule. 64 ≈ 512×512 px on SD1.5's shallowest layer.
• base_global_ratio / peak_global_ratio (0.2 / 0.8) - how much of the global (downsampled whole-image) view each tile gets. Low ratio = more local detail, higher risk of inconsistency; high ratio = safer, less detail.
• tiling_max_depth (2) - how deep into the U-Net the tiling applies.
• slothful (4) - the K/V reduction rate for both input and output blocks, decayed by peak_time / time_decay just like the full node.
• in_blend / out_blend (0) - the pooling blend for input vs output blocks, applied to both K and V together. In the full node these are split into in_k_blend, in_v_blend, out_k_blend, out_v_blend with selectable modes; here one slider does the whole job.

That last collapse is the whole point of the node. If you want the "quieter noise, softer detail" behavior you'd tune in_k_blend for in the full version, you just nudge in_blend up. Keep blends around 0.5 or below; past that, output quality starts paying. And the same warning as the full node applies: a low tile size with a high slothful leaves K and V starved and the image visibly worse.

When to reach for it

When you want the detail-per-tile behavior of Near-sighted Attention but aren't ready to babysit eight style knobs, this is the one. It's also a fine starting point - build a workflow with Simple, figure out which direction you actually want to push the image, then graduate to the full node's separate K/V and AVG/MAX controls if you hit a wall.

Install is the pack's shared story: ComfyUI Manager only knows the pack as "ComfyUI Nearsighted Attention" (the project was renamed after the "Slothful Attention" release, and the old comfyui_slothful_attention URL is a 404 now):

cd ComfyUI/custom_nodes
git clone https://github.com/MitoshiroPJ/comfyui_nearsighted_attention

Restart, and it shows up as "Near-sighted Attention (Simple)". No pip installs, no model files; it only needs what ComfyUI already ships. Tested on SD1.5, SDXL and SSD-1B, fine with LoRA and ControlNet.

Troubleshooting

• Blend too high = mushy output. The fixed 2D-AVG mode means high blends soften outlines fast; back off to 0.3–0.5 if things get blurry.
• Incoherent images mean the tiling is winning: raise peak_global_ratio or base_tile_size before touching slothful.
• Attention patchers (Hypernetwork, ToMe) can conflict - they patch attention too, same caveat as every node in this pack.