Image Saver Metadata

This is the brain to Image Saver Simple's hands. It takes all the details of your generation - prompt, model, sampler, steps, CFG, seed, size - and packages them into a single METADATA object you feed to a saver. On its own it saves nothing. What it does is turn a pile of loose values into the exact A1111-style parameter string Civitai reads for auto-detection, plus the resource hashes that make your uploads link back to the right checkpoint and LoRAs.

Splitting metadata out from file-saving sounds fussy, but it pays off the moment your graph has more than one output branch: build the info once here, route the same METADATA to every Image Saver Simple downstream.

How it works

You wire in the same values your KSampler used - that's the discipline the whole pack is built around, since these nodes can't peek inside your sampler. The node then computes hashes for the model and any LoRAs (including ones you write into the prompt as <lora:name:weight>), formats the A1111 string, and emits it all as a METADATA bundle. A community contribution added the additional_hashes field, where you can paste AutoV3 hashes straight from Civitai for resources the node can't detect on its own - wildcards, ControlNets, anything off-graph - so those get cross-posted too.

Because it computes the metadata but doesn't touch disk, you can also read its string outputs into a preview node to sanity-check what's about to be written.

The inputs and outputs that matter

Everything here is optional - you fill in what you know:

• modelname - the checkpoint name/hash source. Feed it from a loader-with-name node so it matches reality.
• positive / negative - your prompts.
• seed_value / steps / cfg / sampler_name / scheduler_name / denoise - the sampler settings; drive these from the same nodes as your KSampler.
• additional_hashes - paste Civitai AutoV3 hashes for undetectable resources.
• clip_skip - record it if your model family uses it (SD 1.5 anime lineage mostly; on SDXL and newer LLM-encoder models it's usually inert).

The main output is metadata (a METADATA object) - send it to Image Saver Simple. It also emits hashes and a1111_params strings for inspection.

Installing it

ComfyUI Manager: search ComfyUI Image Saver, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/alexopus/ComfyUI-Image-Saver
cd ComfyUI-Image-Saver
pip install -r requirements.txt

Restart, hard-refresh. Small dependency install, no models.

Where people get tripped up

The metadata is only as accurate as what you feed it. Type the model name by hand and get a character wrong, and the hash won't match any Civitai version, so nothing links. The fix is to source modelname from Checkpoint Loader with Name (or UNet loader with Name) so the exact filename flows through untouched.

Watch clip_skip: pass it a string instead of a number and the downstream save can throw bad operand type for abs(): 'str'. Keep it a plain integer. And don't expect resources you never loaded to appear - if a LoRA runs through a node the saver can't inspect, it won't be hashed automatically; that's what additional_hashes is for. Finally, remember this node is inert by itself: no saver downstream means no file and no metadata written.