Metadata Test Sampler

Model loading is slow. Checkpoint, CLIP, VAE - every full run spends its first minutes just pulling weights off disk. When all you want to do is verify that your metadata capture works - that the save node records the right sampler, the right LoRA names, the right hashes - that wait is pure waste. Metadata Test Sampler exists to make that test take two seconds: it fakes being a sampler and emits a solid-color image, but every one of its inputs is real metadata that the save node captures exactly as if you'd run a full pipeline.

It's a testing stub, and it's honest about it. No models load, no GPU sampling happens; you get a batch of flat-colored rectangles with a full metadata payload attached. Wire it into Save Image w/ Metadata Universal, save, and inspect the PNG's metadata. If the parameters are right, your rule config is right - without touching your real models.

The inputs

Everything here is a metadata record disguised as a generation parameter:

• positive_prompt / negative_prompt - recorded verbatim (defaults to a "Synthetic prompt for metadata validation").
• model_name / model_hash, vae_name / vae_hash - fake names and hash stubs (defaults like test_models/fake-model.safetensors, FAKEHASH001). The tooltip is explicit: no file access happens, these are just reported.
• clip_name1 (plus optional clip_name2/3/4) - encoder names recorded in metadata; leave the extras blank to skip.
• seed, steps, cfg, guidance, sampler_name, scheduler - the generation settings the save node normally pulls off a real KSampler. sampler_name defaults to euler, scheduler to normal.
• width / height / batch_size - the size of the synthetic output. Bump batch_size and the saved metadata records batch index/size, which is handy for testing batch handling.
• colour_r / colour_g / colour_b - the RGB of the flat image (each 0–1). Change these if you want distinguishable outputs per run.
• generator_version - an optional version string recorded as "Metadata generator version".

The single images output is a solid-color IMAGE batch. Since batch_size is capped at 8 and dimensions at 4096, it's cheap enough to run dozens of times while iterating on rules.

Actually getting it

The node shows up under SaveImageWithMetaDataUniversal/Test after you install the pack - the |unimeta variant is registered by default. (There's also a plain MetadataTestSampler class, but that one only loads when you set the METADATA_ENABLE_TEST_NODES=1 environment flag before launching ComfyUI - you don't need it, since the |unimeta version is the always-on one.)

The loop it enables

The README's testing advice basically assumes this node exists:

1. Build: Metadata Test Sampler → Save Image w/ Metadata Universal.
2. Save, then drag the PNG back into ComfyUI or open it in a metadata viewer.
3. Missing a field? Re-run the scanner, tweak rules, re-save rules, test again.

Pair it with the METADATA_TEST_MODE=1 env flag (multiline, deterministic parameter output) and you get diffable snapshot output - ideal for confirming that "field X now shows up" without squinting at one giant line. For testing whether your save node is doing its job, nothing beats a node that costs nothing to run.