SMC Make Anchor

SMC Make Anchor (class SMCMakeAnchor) is the pack's explicit anchor factory. It takes your image, shrinks it to a target megapixel count, and hands you two versions: the small anchor itself, and that same small image scaled back up to the original resolution so you can actually look at what the "stable" version of your frame would be.

Why bother? The whole Stable Manifold Compander idea is that a FLUX.2-style model is compositionally more stable at low resolution - 1MP Klein output doesn't have the outward zoom drift, the washout, or the flat chroma that appears when you push to 2-4MP or detail a crop. The anchor is the honest version of the image. If you let SMC Compand run with no anchor_image connected, it derives an anchor internally from config.anchor_mp - so SMCMakeAnchor isn't strictly required. It's for when you want to see the anchor, control it, or pass it to other nodes.

Inputs

• image - your high-res image or crop.
• target_mp (0.1–4, default 1.0) - the megapixel target. 1.0 is the sweet spot for most models' stable regime; the aspect ratio is preserved.
• round_to (1–128, default 16) - the anchor dimensions get rounded to a multiple of this, keeping them friendly to models and VAEs that want multiples of 16.
• upsample_mode - bicubic, bilinear, or nearest-exact. This only affects the anchor_resized output (the small image brought back to full size); the small anchor itself is always bicubic-downscaled.

Outputs

• anchor_small - the actual low-res anchor. This is what you'd feed into the affine workflow as anchor_image.
• anchor_resized - the anchor upscaled back to the original resolution, for inspection or for feeding into SMC Anchor-Guided Compand's anchor_image port (it expects the same resolution as high_image).
• summary - a string like anchor 1024x1024 (~1.048 MP) so you can read the result without wiring up a text node.

Use it this way

The clean pattern for a masked-crop workflow: extract a crop, run your detail pass, then make an anchor from the crop before the detail pass and feed it to SMC Anchor-Guided Compand so the correction has an explicit "how it looked when it was stable" reference. Or just preview anchor_resized to convince yourself that the low-res version really does hold composition better - which is a useful sanity check before you trust any correction.

Install & gotchas

Install via ComfyUI Manager (search "Stable Manifold Compander") or:

cd ComfyUI/custom_nodes
git clone https://github.com/xmarre/ComfyUI-StableManifoldCompander

No extra dependencies, no model downloads. The trap people hit: target_mp is megapixels, not a percentage - set it to 1 for 1 megapixel, not 0.01. And if you're generating the anchor from an image that's already drifted (like the detailed crop rather than the pre-detail source), you're anchoring to the problem, not away from it.