ComfyUI and generative AI

Why connect ComfyUI to Zegami

Generative-AI workflows produce batches, not single images — thousands of variations across seeds, samplers, CFG values, checkpoints, and prompts. Folder-scrolling doesn’t scale. With Zegami you get a visual grid of every output, filter and sort by any generation parameter, cluster by visual similarity (UMAP) to spot duplicates and outliers, and share a public link as the deliverable.

The integration has two halves:

1. The Zegami Batch Export ComfyUI node (open-source, MIT) that captures and uploads each generation.
2. The Zegami side: a collection-scoped API key, and the Calculated Columns → From JSON path tool that turns the captured workflow graph into filterable columns.

1. Create a target collection and a scoped key

1. Create a collection in Zegami (empty is fine — the node will fill it).
2. Open the collection’s Settings → API access tab.
3. Click Generate ComfyUI key. The key is scoped to this collection only — it can’t touch the rest of your account, so it’s safe to keep in a shared workflow file. Copy it now; it’s shown only once.

A scoped key is the recommended choice for ComfyUI. (You can also mint account-wide keys under Account → API keys, but those grant access to every collection you can reach.)

2. Install and configure the node

Install ComfyUI-Zegami from ComfyUI Manager (search “Zegami”), or clone github.com/zegami/comfyui-zegami into ComfyUI/custom_nodes. Video export also needs ffmpeg on your PATH.

Provide your key in priority order (the node never reads it from the workflow JSON, so shared workflows stay credential-free):

1. ZEGAMI_API_KEY environment variable (recommended)
2. ~/.zegami/config.json: { "api_key": "zeg_…" }
3. the node’s api_key_override input

3. Wire up and queue

Drop the Zegami Batch Export node inline anywhere — it passes images through, so it can sit before a Save node.

• Connect your decoded images to the images input (the primary path: each item in the batch becomes a separate collection item).
• For generative video (Wan / Hunyuan / LTX / Mochi), connect frames to the video input; the node encodes an MP4 and a poster thumbnail.
• Set collection_id to your collection’s id.
• Optionally set tags (e.g. a campaign or client name) and notes.

Queue the prompt. Switch to Zegami and your generations appear in the grid as the pipeline processes them. The node uploads on a background queue and never slows or breaks a generation — if Zegami is unreachable, your images are still saved locally.

4. Turn the workflow graph into columns

Every output carries its full ComfyUI prompt graph as an opaque _comfy_json field. To make parameters filterable:

1. Open the collection, go to the views panel and Add calculated column.
2. Switch to the From JSON path mode.
3. Use the Suggested columns chips — Zegami auto-detects common nodes (KSampler, LoraLoader, CheckpointLoaderSimple, CLIPTextEncode) and offers one-click columns for seed, steps, cfg, sampler, scheduler, model, lora, and the positive/negative prompts.
4. Outputs vary per model, so for anything off that list press Discover from data. Zegami samples your actual outputs, walks the JSON, and lists every field it finds — with the inferred type, how often it’s present (fill %), how many distinct values it takes, and an example. Click a row to load it into the form and tweak it, or press + Add to add it directly.
5. Or extract any field by hand: pick the _comfy_json source column, enter a JSON path (e.g. $.prompt.6.inputs.seed), choose a type, and preview.

Mixed workflows? ComfyUI numbers nodes by an internal id that shifts when different graphs feed one collection, so the same field (e.g. seed) can live at a different path in each output. Discovery handles this: it groups by node class and builds a single column that reads the value wherever it lands (a ×N badge marks columns merged across N node-ids), so one seed column works across every model in the collection. A fill % below 100% tells you a field only appears in some outputs.

All of these columns persist with the collection, so they re-apply every time it loads. Now you can filter by sampler, sort by CFG, group by checkpoint, or build a scatter plot of seed vs. any metric.

5. Triage, compare, and publish

• Filter by any extracted parameter to narrow thousands of outputs to the handful worth keeping (Filtering & search).
• Similarity / UMAP surfaces near-duplicates and structural clusters — invaluable for auditing a LoRA or checkpoint training set.
• Video items show a play badge and duration in the grid; hover to preview (where supported), and open the inspector for full playback.
• Publish the curated view as a public, shareable collection (Sharing & publishing) — the gallery itself becomes the client deliverable.

Plan limits

There’s no hard per-collection item cap — your plan’s storage is the real ceiling (shown on Organisation → Plan, with a rough “≈ N images / N video clips” estimate). A studio pushing 10k+ video items should keep an eye on it; Team plans add metered storage beyond the included quota.