comfyui-cyberdelia-metadata

comfyui-cyberdelia-metadata

Civitai-compatible image metadata for ComfyUI, with robust handling of complex conditioning chains and modern multi-sampler workflows.

By Cyberdelia AI Lab · github.com/cyberdeliaAI

What it does

Adds a Cyberdelia save node that writes structured metadata to PNG, JPG, or WebP files. The format is compatible with Civitai, so when you upload your image it reads back your seed, model, LoRAs, prompts, and sampler settings automatically — no manual entry.

What's different in this version

Writing prompt and sampler info into a PNG sounds simple, but breaks down in real workflows. This release is focused on the rough edges:

Version 2.1.0 — CyberKrea metadata

• Adds native metadata extraction for CyberKreaSampler and CyberKreaEmptyLatent without requiring an extra save node or new workflow connection.
• Records the real seed, step count, CyberKrea sampler/schedule, base CFG, generation dimensions, preset, guidance mode, restart settings, detail, eta, sigma gate, and manifold contraction.
• Recovers both original NegPiP prompt strings when CyberKrea's optional negative input is deliberately left disconnected.
• Prevents a seed relayed through rgthree's steps channel from replacing an already detected sampler step count.

Version 2.0.1 fixes

• Restores node previews by returning output-relative, URL-safe subfolder paths to ComfyUI.
• Automatically recognizes workflows installed through the Registry, Git, ComfyUI-Manager custom install, and the tagged revived predecessor pack when migrating legacy node ids.
• Keeps ComfyUI's validated safe basename when subdirectory_name is set. In that mode, directory components in filename_prefix are intentionally not reused.

Correct prompt/sampler attribution in complex graphs

• ConditioningZeroOut handling — when your negative is zeroed out, metadata reports it as empty instead of walking past the zero-out and picking up whatever CLIPTextEncode was upstream.
• rgthree Context Big / Context Switch — the walker follows the right input slot through these passthrough nodes, so the negative branch doesn't accidentally pull text from the positive branch.
• ControlNet apply chains — passthrough resolution that respects positive/negative separation.
• Multi-sampler workflows — in base + upscale pass setups, the primary generation sampler (farthest from the save node) is reported, not whichever was found first.

Runtime text capture

Nodes that compute their final text at runtime — wildcard expanders, dynamic prompts, LLM-based prompt engineers like Cyberdelia Z-Engineer — can register their resolved text via the hook API and have it captured in metadata, instead of falling back to a raw widget value or unresolved wildcard placeholder.

Broad third-party integration

Out-of-the-box support for ~20 popular custom node packs (rgthree, efficiency-nodes, easyuse-nodes, lora-manager, RES4LYF, WanVideoWrapper, and more — see modules/defs/ext/). Adding a new one is a small Python file following an existing pattern.

Output flexibility

PNG (lossless), JPG, or WebP (lossy or lossless) at adjustable quality. Optional sidecar .json workflow file. Subdirectory templating with date masks, model name, and prompt prefixes. Five metadata scopes from full to none.

Installation

Via ComfyUI-Manager

Search for comfyui-cyberdelia-metadata or Cyberdelia and install.

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/cyberdeliaAI/comfyui-cyberdelia-metadata.git

Restart ComfyUI.

Version 2 uses installation-wide unique serialized ids. It can coexist with nkchocoai/ComfyUI-SaveImageWithMetaData; neither pack overwrites the other's node registration.

Usage

Replace your Save Image node with Save Image With Metadata (Cyberdelia). Hook up the image input — the rest is automatic.

If you use rgthree, you can optionally connect the CONTEXT output of Context or Context Big to the node's context input. Non-empty context values for seed, steps, CFG, checkpoint, sampler, and scheduler take priority over automatic graph detection. The context conditioning selects the prompt; its text fields and CLIP dimensions fill gaps only when prompt text or generation size cannot otherwise be detected. Empty context fields keep using the normal graph fallback, so the input is safe to leave disconnected and existing workflows remain unchanged.

Node options

| Option | Description | |---|---| | context | Optional rgthree CONTEXT; populated values are authoritative metadata overrides | | filename_prefix | Filename template — supports mask tokens (see below) | | subdirectory_name | Relative subdirectory inside ComfyUI's output folder — supports mask tokens | | output_format | png, jpg, webp, or any of those + _with_json for a sidecar workflow file | | quality | max / lossless WebP (100%), high (80%), medium (60%), low (30%). PNG ignores this. | | metadata_scope | full (default + extras), default (Comfy stock), parameters_only (A1111 string), workflow_only, or none |

Filename and subdirectory templating

filename_prefix and subdirectory_name accept these mask tokens:

| Token | Replacement | |---|---| | %seed% | Seed value | | %width% / %height% | Image dimensions | | %pprompt% / %pprompt:[n]% | Positive prompt (optionally first n characters) | | %nprompt% / %nprompt:[n]% | Negative prompt (optionally first n characters) | | %model% / %model:[n]% | Checkpoint name (optionally first n characters) | | %date% | Date as yyyyMMddhhmmss | | %date:[format]% | Date in a custom format |

Date format identifiers: yyyy (year), MM (month), dd (day), hh (hour), mm (minute), ss (second). Example: %date:yyyy-MM% produces 2026-05.

When subdirectory_name is empty, directory components in filename_prefix continue to create subdirectories as supported by ComfyUI. When subdirectory_name is set explicitly, it alone selects the destination directory and the filename keeps the basename validated by ComfyUI.

Runtime text capture (for node authors)

Building a custom node that computes its final prompt text at runtime? You can register that text so it ends up in saved metadata instead of whatever the user typed in the widget. The mechanism is loose-coupled — no hard import dependency on this extension:

import sys
from comfy_execution.utils import get_executing_context

context = get_executing_context()
if context is not None:
    for mod in sys.modules.values():
        if mod is None:
            continue
        record_fn = getattr(mod, "record_resolved_text", None)
        if callable(record_fn):
            record_fn(context.node_id, final_text, getattr(context, "list_index", None))

If users don't have a compatible metadata extension installed, the snippet does nothing. This is how Cyberdelia Z-Engineer gets its LLM-engineered output into metadata.

Supported third-party nodes

Each file in modules/defs/ext/ registers a third-party node pack. Currently covered: CyberKrea Sampler, rgthree, efficiency-nodes, easyuse-nodes, lora-manager, RES4LYF, WanVideoWrapper, Lightx02-Nodes, comfyui-custom-scripts, comfyui-clip-with-break, comfyui-easy-civitai-xt-nodes, comfyui-flux-settings-node, comfyui-gguf, comfyui-miaoshouai-tagger, comfyui-restart-sampling, comfyui-weilinnodes, ComfyUI-NegPiP-ZImage, CheckpointDiscoveryHub, CR_ApplyLoRAStack, everywhere, size_from_presets, SantodanNodes.

[!TIP] If the full metadata scope errors out, it's usually an unrecognised third-party node in your workflow. Either swap to a Comfy Core equivalent or add a new file under modules/defs/ext/ following the existing pattern.

Migrating workflows to version 2

The legacy ids SaveImageWithMetaData and CreateExtraMetaData are owned by the original nkchocoai pack as well. Version 2 therefore uses the unique ids CyberdeliaSaveImageWithMetaData and CyberdeliaCreateExtraMetaData. Registering the old ids as aliases would reintroduce the same load-order conflict, so they are deliberately not exposed by the backend.

Normal UI workflows created with this fork are migrated automatically before ComfyUI checks for missing nodes when their cnr_id or aux_id identifies the Registry package, the cyberdeliaAI/comfyui-cyberdelia-metadata Git repository, or the tagged revived predecessor package. Save the workflow once to persist the new ids. Workflows belonging to nkchocoai are left untouched.

For a directory of workflow JSON files, first run the bundled migrator in dry run mode:

python3 scripts/migrate_workflows.py /path/to/workflows

Then write the changes; each changed file gets a .json.bak backup:

python3 scripts/migrate_workflows.py --write /path/to/workflows

Old API JSON and workflows without a recognized cnr_id or aux_id cannot be distinguished from the original pack automatically. If the selected files are definitely meant to use the Cyberdelia nodes, opt in explicitly:

python3 scripts/migrate_workflows.py --all-legacy --write /path/to/workflows

This package remains the direct successor to revived_comfyui_image_metadata_extension. Tagged predecessor workflows are migrated automatically; untagged ones need the same one-time --all-legacy migration because they also use the shared legacy ids.

Credits

Built on the work of:

• edelvarden — comfyui_image_metadata_extension — original concept and initial implementation.
• Santodan — revived_comfyui_image_metadata_extension — picked up maintenance and added LoRA metadata, subdirectory templating, format/quality controls, and metadata scope options.
• Cyberdelia AI Lab — this version — conditioning chain resolution rewrite, multi-sampler workflow support, runtime text capture, and ongoing maintenance.

License

GPL-3.0 — see LICENSE. Inherited through the fork chain; derivative works must remain GPL-3.0 and preserve copyright notices of all prior authors.