Nougan MOSS RefMod Encode 🎯

Nougan MOSS RefMod Encode 🎯 is a drop-in for Text Encode MOSS-Image Ostris Edit. Same outputs, same conditioning shape, one difference: the references come from baked files in models/refmods/moss rather than from a row of Load Image nodes you keep re-pointing.

That's it. That's the whole value proposition, and if you run the same character through twenty prompts a week it's worth the swap - you stop re-encoding latents and you stop hunting for that one good reference folder.

What goes in

clip - your MOSS-Image text encoder. prompt and negative_prompt as usual.

refmod_1 (required, None allowed), refmod_2, refmod_3 - dropdowns from the MOSS bake library. Up to three per node; for more, chain another encode or bake fewer.

extra_image + vae - an unbaked reference, encoded on the fly. extra_image_scale gives it the same two options the Bake node has:

• kontext (matches Ostris example) - FluxKontextImageScale's nearest ~1MP preset by aspect, center cropped. The convention training used.
• megapixels - whole image, scaled to extra_image_resolution² pixels. Nothing cropped.

extra_image_resolution (default 1024, 0 = keep original size) applies to the megapixels path.

ref_strength - and this is the one worth understanding. It's a schedule, not an amplitude: the node builds conditioning with the references and conditioning without, and uses the reference version for the first ref_strength fraction of the denoising steps. 1.0 runs them every step, 0.6 hands the last 40% to the prompt alone, 0 disables them. Lower means looser likeness and more freedom for your words. If MOSS is clinging to your reference's composition when you wanted a new pose, shorten the reference's shift instead of trying to outweigh it in the prompt.

What comes out

positive and negative (CONDITIONING). Plural references go onto both, matching what the Ostris encode node does, so a workflow swap is a literal one-for-one.

The prerequisite that decides whether any of this works

Stock MOSS-Image ignores reference latents entirely. The conditioning this node emits only does something when:

• your MOSS-Image model or LoRA is edit-trained (ai-toolkit's edit training path), and
• ostris/ComfyUI-MOSS-Image-Ostris-Edit is installed and its MOSS-Image Ostris Edit Model Patch node is on the MODEL line.

Without the patch, the node runs, the graph completes, and the character is absent. That's the classic silent failure in this corner of the ecosystem, so check it before you debug anything else.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/Nougan_Nodes
git clone https://github.com/ostris/ComfyUI-MOSS-Image-Ostris-Edit
# restart ComfyUI

Both are on ComfyUI Manager. The pack installs no pip dependencies of its own. Restart, hard-refresh the browser for the refmod picker, and make sure at least one MOSS character exists - otherwise the dropdown offers nothing but None.

Where people get burned

Empty dropdown. Nothing baked yet, files in the wrong folder, or they're in another family's subfolder (krea2, qwen21, flux3). The MOSS list only shows MOSS refmods; each family carries its own metadata key and they don't cross over.

Scale mismatch between bake and encode. If references were baked in kontext mode and your extra image is going through megapixels, you're mixing conventions in one conditioning. Pick one and stay there.

Token creep. Three references at ~1MP each is a lot of tokens in every block of every step. MOSS is not a cheap model to stack references on; two is a sensible default.

Assuming more references means more likeness. Past two or three, you're usually adding contradictions, not fidelity. Character consistency work is a "less but better references" discipline - one clear front-facing portrait beats six near-duplicates.