comfyui-save-image-rich-metadata

ComfyUI — Save Image (Rich Metadata)

A drop-in replacement for ComfyUI's SaveImage that writes rich, multi-format metadata into every PNG — fully compatible with A1111 / CivitAI, plus a clean authoritative JSON chunk consumed by the AI Gallery app.

Built on the ComfyUI v3 API with Autogrow — unlimited image input slots (framework cap: 100).

What it writes

Each PNG gets three tEXt chunks:

| Chunk | Format | Purpose | |---|---|---| | ai_gallery_meta | JSON | Canonical, authoritative metadata (consumed by AI Gallery, no heuristics needed) | | prompt + workflow | JSON | Standard ComfyUI — drag the image back to ComfyUI to restore the workflow | | parameters | A1111 text | CivitAI, stable-diffusion-webui, A1111 ecosystem |

CivitAI compatibility

The parameters chunk follows the A1111 webUI format exactly, including resource hashes so CivitAI auto-detects and links the checkpoint and LoRAs:

<positive prompt> <lora:name_1:weight_1> <lora:name_2:weight_2> …
Negative prompt: <negative prompt>
Steps: N, Sampler: name, CFG scale: X, Seed: N, Size: WxH, Model hash: abc123def456, Model: name, Lora hashes: "name_1: 0011aabbccdd, name_2: …", Hashes: {"model": "abc123def456", "lora:name_1": "0011aabbccdd"}

This is the canonical format CivitAI's PNG inspector parses on upload — drop any image saved by this node onto CivitAI and the positive prompt, negative prompt, model, sampler, steps, CFG, seed, dimensions and inline LoRAs are extracted automatically, and the model/LoRA resource pages are linked via their hashes.

Hashes are the AutoV2 form (first 12 hex of the file's SHA256), computed for the checkpoint/UNet (searched in checkpoints → diffusion_models → unet) and every LoRA (loras). Each file is hashed once and cached in .hash_cache.json (keyed by path + size + mtime), so the first save after loading a new model is slower and subsequent saves are instant. Nothing is written next to your model files.

Run-time prompts: prompt_text / negative_text inputs

The workflow graph only stores node inputs. When your prompt is produced while the workflow runs — an LLM prompt expander (TextGenerate, Ollama, Florence/Qwen captioners), a wildcard processor, a random picker — the final text is not in the graph and no parser can recover it after the fact.

For those workflows connect the actual STRING to the node's optional prompt_text input (and negative_text if you have one). Whatever you plug in there is written verbatim to ai_gallery_meta.prompt and to the A1111 parameters chunk, overriding graph extraction. Leave the inputs unconnected for ordinary static-prompt workflows.

What it extracts from the workflow

Walks the execution graph (no keyword guessing):

• prompt / negative — traces the sampler's positive / negative link (or its guider, incl. BasicGuider.conditioning) back to the text through CLIPTextEncode, CLIPTextEncodeFlux (t5xxl/clip_l), string primitives (value / string / prompt, e.g. CR Prompt Text, String Literal), PreviewAny, boolean routers (ComfySwitchNode, Crystools Switch any, rgthree Any Switch — switch literal or linked to a PrimitiveBoolean, with fallback to the other branch), Text Concatenate / StringConcatenate (joined with the node's delimiter), SDXLPromptStyler, ImpactWildcardEncode, Qwen image-edit encoders, and conditioning pass-throughs with (positive, negative) in and out (LTXVConditioning, WanImageToVideo — the output slot picks the side). LLM/VLM generator nodes are recognised as unrecoverable: if a ShowText node displays their output its cached text is used, otherwise the router's other branch (your raw prompt) is taken — or use prompt_text above. ConditioningZeroOut → empty negative; a negative identical to the positive (one Flux encoder wired to both) is dropped.
• sampler, steps, cfg, seed — from KSampler* / SamplerCustom*, or the RandomNoise / BasicScheduler / *Guider / KSamplerSelect helpers of SamplerCustomAdvanced graphs. Values wired from primitive/seed nodes (Seed (rgthree), PrimitiveFloat, Seed Generator, …) are followed.
• model_name — from CheckpointLoader* / UNETLoader* / UnetLoader*
• loras — only enabled entries, de-duplicated, from:
  • stock LoraLoader, LoraLoaderModelOnly
  • rgthree Power Lora Loader (dict slots, respects on: false)
  • rgthree Lora Loader Stack (lora_01 + strength_01, skips None)
  • CR LoRA Stack and other lora_name_N stacks (respects switch_N)
  • LoraLoaderStackedAdvanced (lora_name widget dict + lora_weight)

Custom samplers & runtime prompt builders

The graph walk also handles modern Flux / SD3-style pipelines where the prompt isn't a plain static string on the sampler:

• Guider-based custom samplers — SamplerCustomAdvanced and friends route conditioning through a guider node (BasicGuider, CFGGuider, DualModelGuider, …) instead of exposing positive/negative directly. We follow the guider link to recover both, and read cfg from it.
• Ideogram 4 Prompt Builder (KJNodes) — this node computes its caption JSON at runtime, so there's no static text anywhere in the graph. We reproduce the builder's exact assembly from its inputs (high_level_description, background, style_description, elements, color palettes) to recover the real prompt that conditioned the image.
• ConditioningZeroOut — treated as an empty negative, so a zeroed-out negative branch never echoes the positive prompt it wraps.

Screenshots

Full workflow

Pipeline tail (KSampler → VAEDecode → Save Image)

Install

cd ComfyUI/custom_nodes
git clone https://github.com/quzopl/comfyui-save-image-rich-metadata.git

Restart ComfyUI. The node appears as Save Image (Rich Metadata) in the image category.

Requires ComfyUI with the v3 API (comfy_api.latest) — present in all recent releases.

Usage

Replace SaveImage in your workflow with Save Image (Rich Metadata). Connect the images input, set filename_prefix (e.g., MyRender).

Unlimited image inputs

The images input is an autogrow slot — the node UI adds a new empty slot whenever you connect one. Plug in as many independent image batches as you want (e.g., raw + refine + upscale + variations all in one workflow).

Each connected batch is saved separately:

• Slot 1 → filename_prefix
• Slot 2 → filename_prefix_2
• Slot 3 → filename_prefix_3
• … and so on

All slots share the same workflow metadata (same ai_gallery_meta, prompt/workflow, and parameters chunks) since they're produced by the same execution graph.

Optional flags

• embed_workflow (default ON) — also embed standard ComfyUI chunks
• embed_a1111 (default ON) — also embed A1111-compatible parameters (turn off only if you need a minimal PNG for some reason)

Dependencies

None. Uses only Pillow and NumPy (both already shipped with ComfyUI).

License

MIT