Which base model makes your LoRA look best? Drop your MOSS models in a folder, queue once per model, and get one labeled grid. Every model renders with the steps, CFG, sampler and scheduler its author recommends, read from its Civitai page. That's the method from my article Which base model is best for your LoRA? (part 2 of my LoRA guide).
What it does
Checkpoint Folder Iterator (
model_type: diffusion_model) loads one UNET per queued run from a folder inmodels/diffusion_models(defaultMoss). Only one model is in memory at a time, so the folder can be as big as you like.Model Recommended Settings reads each model's recommended settings from the Civitai info that LoRA Manager (or Civitai Helper) already saved. It reads the settings the author calls recommended first, then the most common values on the model's example images. Settings used shows what each run got and where it came from. Nothing is downloaded.
Turn the
use_model_*switches off to give every model the same settings instead.A scheduler that isn't installed (e.g.
bong_tangent) falls back to the widget, or to a substitute you set inComfyUI/user/kikotools/model_settings.json:{"schedulers": {"bong_tangent": "beta"}}. Per-model overrides go in the same file.
The LoRA stack applies to every model, with its trigger words put in front of the prompt. It's set up for my mbl9 LoRA. Switch the LoRA off to compare the bare models.
Model Compare Sheet collects the tiles into
grid.png, labeled with the file names. It starts a new grid by itself when you change the prompt or settings mid-sweep, and the old tiles move to a timestamped folder.Prompt Manager text node, Kiko Prompt Enhancer (off), Model Resolution Picker (pick 1088×1920 for portraits).
How to run
Set folder (and optionally include/exclude regex) on Model folder. Only
.safetensorsfiles are picked up.Leave index at 0 with increment.
Set the Run button's batch count to the number of models (Models in folder shows it after the first run) and queue.
A run past the last model stops with "past the end" instead of starting over. Set index back to 0 for a new sweep.
Your models need their Civitai info. In LoRA Manager, open the Checkpoints page (it lists diffusion models too) and fetch the metadata for each model. Models it can't find just use the widget values.
Files
kiko-moss-model-sweep-v7.json: the workflow.kiko-moss-model-sweep-workflow.png: a picture of the whole graph, with the workflow embedded.Example grid: mbl9 at 1.0 on 8 MOSS models, the default prompt, 1088×1920, seed 1010, each model on its own recommended settings.
Custom nodes
ComfyUI-KikoTools 1.7.0+ (Checkpoint Folder Iterator, Model Recommended Settings, Model Compare Sheet)
ComfyUI-Lora-Manager (also the source of the Civitai info)
Installing KikoTools: the Comfy Registry lists KikoTools as flagged, so ComfyUI Manager 4.2 (the version ComfyUI ships) won't install it. Either update the Manager to 4.3 or newer and set allow_flagged_nodepack_install = true in ComfyUI/user/__manager/config.ini, or install from GitHub: git clone https://github.com/ComfyAssets/ComfyUI-KikoTools in ComfyUI/custom_nodes, then pip install -r requirements.txt with ComfyUI's Python, and restart.
Models
Your MOSS models in
models/diffusion_models/Moss/. The official ones are MOSS Image Base and SFT. The article lists the seven finetunes I swept.qwen3.5_9b_w4a8_gptq.safetensors→models/text_encoders/: https://huggingface.co/Comfy-Org/MOSS-Image/resolve/main/text_encoders/qwen3.5_9b_w4a8_gptq.safetensorsFlux.2-D/flux2-vae.safetensors→models/vae/: https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy/resolve/main/vae/flux2-vae.safetensorsmbl9 LoRA →
models/loras/, or your own.
The bypassed Download models group fetches the text encoder and VAE. Turn it on in Toggle stages, queue once, then turn it off again.
Based on HaDeS's "MOSS Suite – Compare Models" idea, rebuilt with core and KikoTools nodes.
Happy comparing! Kiko
Description
First public release. Needs ComfyUI-KikoTools 1.7.0+ (Checkpoint Folder Iterator, Model Recommended Settings, Model Compare Sheet). Set up for the mbl9 LoRA with the Comfy-Org qwen3.5_9b_w4a8_gptq text encoder.

