Model Analyzer (Checkpoint)

Everyone who's ever downloaded a checkpoint has asked the same question: what CFG and sampler does everyone else actually run this at? The model card usually gives you nothing, or one screenshot's worth of settings that may or may not be typical. Model Analyzer (Checkpoint) answers that properly - it pulls a real sample of the community's showcase images for your checkpoint and boils them down into an actual report instead of a guess.

This is the pack's "In-Depth Model Analysis" feature, and per the README it's meant to replace an older, clunkier multi-node version of the same idea with one node that does fetch → tally → summarize → markdown in a single pass.

How it works

It hash-matches your local checkpoint against Civitai, pulls a batch of the community images generated with it, and counts what actually shows up: which sampler, which CFG range, which steps, and which LoRAs get paired with it most often - the "golden combination" discovery the README talks about. That gets written out as a markdown report, plus a machine-readable bundle for downstream nodes.

The inputs worth knowing

Two shape the sample you're pulling from, the rest are size/filter knobs:

• sort - Most Reactions, Most Comments, or Newest. This decides which images you're sampling, and it matters: Most Reactions skews toward all-time favorites, Newest skews toward whatever's trending this week. Don't be surprised if the two give slightly different "common" settings.
• image_limit (1–1000, default 100) - how many sample images to crunch. More gives you a steadier statistical picture; it's also more API calls, so don't max it out on a whim.
• nsfw_level (None / Soft / Mature / X) and filter_type (all / image / video) narrow what gets pulled in.
• summary_top_n (1–100, default 10) - how many top prompts/combos actually get listed in the report.
• force_refresh - same deal as everywhere else in this pack: no reuses the cached fetch, yes forces a fresh pull.

Outputs

• full_report_md - the whole thing, meant for the Markdown Presenter node so it renders instead of sitting as a wall of text.
• fetch_summary - a one-liner like "Analyzed 100 entries successfully," useful for a quick sanity check without opening the full report.
• params_pipe (CIVITAI_PARAMS) - the tallied "community common" settings, packaged for the companion node, Get Parameters from Analysis, to unpack into individual sampler-ready values.

Installing it

Same pack, same two routes: search Civitai Toolkit in ComfyUI Manager, or

cd ComfyUI/custom_nodes
git clone https://github.com/BAIKEMARK/ComfyUI-Civitai-Toolkit
pip install -r requirements.txt

Restart ComfyUI, then look under Civitai → 📊 Analyzer.

The honest caveat

Treat what this node hands you as a strong starting point, not gospel. "Common" CFG and sampler are architecture-dependent - SD1.5/SDXL checkpoints cluster around CFG 5–9 with a DPM++/Karras-style sampler, while distilled and flow-matching checkpoints (Turbo, Lightning, anything Flux-lineage) want CFG closer to 1 and a completely different sampler family. If the report's numbers look weirdly aggressive or oddly flat for what you know the architecture, check what the model card says it was built on before trusting the crowd average over it.

Troubleshooting

The report comes back thin even with image_limit maxed out. That usually means the checkpoint just doesn't have a big enough public gallery on Civitai - a niche or recent merge with a handful of showcase images will always produce a short report. Not a fetch failure, just a small sample.

Nothing comes back at all, or fetch_summary reports zero. Two likely causes: the hash didn't match anything on Civitai (a private finetune, a heavily repacked file, or a checkpoint that's since been pulled - model removals happen on Civitai on a regular basis, not as a rare event), or you're getting rate-limited. If it's the latter, add a Civitai API key in the toolkit's settings - the pack added dedicated support for this specifically to raise request limits.

Changed a filter and the report looks unchanged. Set force_refresh to yes. The node caches aggressively for speed, and a stale cache is the first thing to rule out before assuming something's broken.