ComfyUI-Info-Prompt-Toolkit

English | 日本語

ComfyUI-Info-Prompt-Toolkit

Selectively encrypt prompts, model details, and other generation data you do not want to expose, while saving the rest as A1111/Civitai-compatible metadata. This toolkit’s reader nodes restore encrypted fields, so you can control what an image reveals without giving up reusable generation settings.

Built around image_info, it simplifies ComfyUI wiring and makes trial results easier to carry into your next production pass. It also includes same-name .txt caption saving, XY Plot, Tiled Sampling (SDXL (with ControlNet Tile) and Anima), SAM3, Detailer, PixAI Tagger, OppaiOracle Tagger, wildcards, and Dynamic Prompts.

Guide: Using A1111/Civitai-Compatible Metadata: Encrypt Only the Generation Data You Don’t Want to Share — An illustrated guide to saving, restoring, reusing, and selectively encrypting image metadata.

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/kinorax/comfyui-info-prompt-toolkit.git
cd comfyui-info-prompt-toolkit
pip install -r requirements.txt

Installing via ComfyUI Manager is the easiest option.

Note: SAM3 Prompt To Mask, PixAI Tagger, and OppaiOracle Tagger require manual placement of model files under ComfyUI/models after installation. See the relevant sections in Core Nodes for details.

Support Ongoing Development

A paid note article is available for users who want polished results with less trial and error. It introduces production-oriented SDXL and Anima workflows built with these custom nodes. Purchasing the article helps support ongoing development.

Article: ComfyUI SDXL & Anima Workflow: Automatic Detailer, XY Plot, Caption File Generation, and More

The workflow files use English labels throughout.

Core Nodes

SamplerCustom (Sampler Params, Tiled) / SamplerCustom (Sampler Params)

<a href="assets/readme/samplercustom-sampler-params-pair.webp"><img align="left" hspace="16" src="assets/readme/samplercustom-sampler-params-pair.webp" alt="SamplerCustom pair" width="210"></a>

<ul> <li>Accepts <code>sampler_params</code> and performs sampling using the official <code>KSamplerSelect</code> / <code>BasicScheduler</code> / <code>SamplerCustom</code> nodes (no custom sampling logic is implemented).</li> <li>The Tiled version uses the same sampling settings as the standard version and runs inference by splitting into spatial tiles.</li> <li>The Tiled version is useful not only under VRAM constraints, but also when you need output resolutions beyond a model's practical supported size.</li> <li>Models with verified operation for the Tiled version are currently <code>SDXL (with ControlNet Tile)</code> and <code>Anima</code>. Other models may still work, but compatibility is not guaranteed.</li> </ul> <br clear="left">

Sampler Params

<a href="assets/readme/sampler-params.webp"><img align="left" hspace="16" src="assets/readme/sampler-params.webp" alt="Sampler Params" width="210"></a>

<ul> <li>Aggregates sampler/scheduler/steps/denoise/seed/cfg into a single <code>sampler_params</code> bundle.</li> <li>Improves both wiring simplicity and configuration reuse.</li> </ul> <br clear="left">

XY Plot Start / XY Plot Modifier / XY Plot End

<a href="assets/readme/xy-plot-start-modifier-end.webp"><img align="left" hspace="16" src="assets/readme/xy-plot-start-modifier-end.webp" alt="XY Plot nodes" width="210"></a>

<ul> <li><code>XY Plot Modifier</code> chains axis-specific change sets (<code>label</code> + <code>changes</code>) as an array and defines X/Y condition variations.</li> <li><code>XY Plot Start</code> applies X/Y modifiers to the base <code>image_info</code>, generates per-cell <code>image_info</code>, and outputs in <code>list</code> (all at once) or <code>loop</code> (iterative) mode.</li> <li><code>XY Plot End</code> receives <code>xy_plot_info</code> and generated images, then composes the axis-labeled grid image <code>xy_plot_image</code>.</li> <li>In <code>loop</code> mode, <code>Start</code> and <code>End</code> cooperate via <code>loop_control</code> to process all cells sequentially.</li> </ul> <br clear="left">

Image Info Context

<a href="assets/readme/image-info-context.webp"><img align="left" hspace="16" src="assets/readme/image-info-context.webp" alt="Image Info Context" width="210"></a>

<ul> <li><code>Image Info Context</code> reconstructs updated <code>image_info</code> by reflecting only the input fields that are actually connected.</li> <li>When <code>positive</code> is connected, it extracts <code>&lt;lora:...:...&gt;</code> tags and applies the tag-removed <code>positive</code> text and extracted values to <code>lora_stack</code>.</li> <li><code>extras</code> is merged into existing <code>image_info.extras</code>, and duplicate keys are overwritten by input-side values.</li> <li>In addition to updated fields, it outputs <code>base_sampler_params</code> (with fixed <code>denoise=1.0</code>) for easy connection to downstream sampling nodes.</li> </ul> <br clear="left">

Image Info Fallback

<a href="assets/readme/image-info-fallback.webp"><img align="left" hspace="16" src="assets/readme/image-info-fallback.webp" alt="Image Info Fallback" width="210"></a>

<ul> <li><code>Image Info Fallback</code> fills missing fields in the primary <code>image_info</code> from <code>image_info_fallback</code>.</li> <li>Fallback is limited to missing fields, and fields that already have values in the primary side are not overwritten.</li> <li><code>extras</code> is merged by adding only missing keys, while existing keys are preserved.</li> <li><code>lora_stack</code> is supplemented when it is missing or <code>None</code>; an explicit empty list means no LoRA and is preserved.</li> </ul> <br clear="left">

Detailer Start / Detailer End

<a href="assets/readme/detailer-start-end.webp"><img align="left" hspace="16" src="assets/readme/detailer-start-end.webp" alt="Detailer Start and End" width="210"></a>

<ul> <li><code>Detailer Start</code> detects a <code>bounding box (bbox)</code> from regions where the input <code>mask</code> value is greater than <code>0.0</code>, then creates an expanded processing region using <code>crop_margin_scale</code>.</li> <li>The processing region is pre-upscaled by <code>upscale_factor</code>, then rounded up to a size divisible by <code>mini_unit</code> (8/16/32/64) and arranged with <code>center padding</code>.</li> <li>For compositing back into the original image, the graded values of the input <code>mask</code> are preserved as a <code>soft mask</code>, and that information is stored in <code>detailer_control</code>.</li> <li><code>Detailer End</code> receives <code>inpainted_image</code> and <code>composite</code>s it back into the original image using the preserved <code>soft mask</code>.</li> <li>If <code>inpainted_image</code> resolution differs from the expected <code>canvas</code> size, it is automatically resized before restoration.</li> <li>If no target region is detected, <code>Detailer End</code> does not evaluate <code>inpainted_image</code> and returns the original image unchanged.</li> </ul> <br clear="left">

SAM3 Prompt To Mask

<a href="assets/readme/sam3-prompt-to-mask.webp"><img align="left" hspace="16" src="assets/readme/sam3-prompt-to-mask.webp" alt="SAM3 Prompt To Mask" width="210"></a>

<ul> <li><code>SAM3 Prompt To Mask</code> is a single-purpose node that generates a <code>soft mask</code> from text prompts. Comma-separated <code>prompt</code> tokens are treated as OR conditions, and accepted results can be controlled with <code>score_threshold</code> and <code>combine_mode</code> (<code>union</code> / <code>top1</code>).</li> <li>The output is a non-binarized <code>soft mask</code>; thresholding and region cleanup are intended to be handled by downstream nodes.</li> <li>This node is designed with a strong focus on keeping additional <code>pip</code>-installed dependencies to a minimum, because some existing SAM3-related extensions can significantly alter Python environments.</li> <li>Under that policy, SAM3-related functionality is intentionally limited to this single purpose, and no further feature additions are planned at this time.</li> <li>Place <code>sam3.pt</code> in <code>ComfyUI/models/sam3</code> (download: <a href="https://huggingface.co/facebook/sam3">facebook/sam3</a>; operation with SAM3.1 and later is unverified).</li> </ul> <br clear="left">

PixAI Tagger

<a href="assets/readme/pixai-tagger.webp"><img align="left" hspace="16" src="assets/readme/pixai-tagger.webp" alt="PixAI Tagger" width="210"></a>

<ul> <li><code>PixAI Tagger</code> estimates tags from images and outputs <code>general</code> / <code>character</code> / <code>ip</code>.</li> <li>Tag selection uses <code>threshold_general</code> / <code>threshold_character</code> and <code>max_tags_general</code> / <code>max_tags_character</code>.</li> <li><code>sort_order</code> controls the final display order for <code>general</code> tags only (<code>score</code> / <code>tag_id</code>); <code>character</code> and <code>ip</code> are always output in score order.</li> <li>This node runs on a PyTorch implementation and does not depend on <code>onnxruntime</code> / <code>onnxruntime-gpu</code>, making the setup less susceptible to Python environment changes.</li> <li>Place the PixAI Tagger v0.9 bundle (<code>model_v0.9.pth</code> / <code>tags_v0.9_13k.json</code> / <code>char_ip_map.json</code>) in <code>ComfyUI/models/pixai_tagger</code> or <code>ComfyUI/models/pixai_tagger/v0.9</code> (download: <a href="https://huggingface.co/pixai-labs/pixai-tagger-v0.9">pixai-labs/pixai-tagger-v0.9</a>).</li> <li>The previous released node remains available as <code>PixAI Tagger (Deprecated)</code> for existing workflows.</li> </ul> <br clear="left">

OppaiOracle Tagger

<a href="assets/readme/oppai-oracle-tagger.webp"><img align="left" hspace="16" src="assets/readme/oppai-oracle-tagger.webp" alt="OppaiOracle Tagger" width="210"></a>

<ul> <li><code>OppaiOracle Tagger</code> estimates tags from images and outputs <code>tags</code> / <code>rating</code>.</li> <li>Tag selection uses <code>threshold</code> and <code>max_tags</code>; <code>sort_order</code> controls only the final display order (<code>score</code> / <code>tag_id</code>).</li> <li>This node runs on a PyTorch implementation and does not depend on <code>onnxruntime</code> / <code>onnxruntime-gpu</code>, making the setup less susceptible to Python environment changes.</li> <li>Place the OppaiOracle bundle (<code>model.safetensors</code> / <code>config.json</code> / <code>preprocessing.json</code> / <code>vocabulary.json</code>) in <code>ComfyUI/models/oppai_oracle/V1.1</code> or <code>ComfyUI/models/oppai_oracle/V1.0</code> (download: <a href="https://huggingface.co/Grio43/OppaiOracle">Grio43/OppaiOracle</a>).</li> </ul> <br clear="left">

Flatten Prompt for Caption / Merge Caption Tokens / Remove Caption Tokens

<a href="assets/readme/caption-nodes.webp"><img align="left" hspace="16" src="assets/readme/caption-nodes.webp" alt="Caption nodes" width="210"></a>

<ul> <li><code>Flatten Prompt for Caption</code> prepares prompt text for caption files by removing prompt emphasis syntax and weight values, then unescaping escaped parentheses and backslashes.</li> <li><code>Merge Caption Tokens</code> merges <code>start</code> and <code>end</code> caption token strings, removes duplicate tokens, and rebuilds the result as a single-line caption.</li> <li>Caption tokens are split on <code>, </code> and <code>. </code>. Existing order is preserved, and the first occurrence of each token is kept.</li> <li><code>Remove Caption Tokens</code> removes exact token matches from <code>string</code> using the comma-separated <code>remove</code> input, while preserving the remaining token order.</li> <li>These nodes are useful when combining tagger output, handwritten captions, and prompt text before writing caption files.</li> </ul> <br clear="left">

Prompt Template

<a href="assets/readme/prompt-template.webp"><img align="left" hspace="16" src="assets/readme/prompt-template.webp" alt="Prompt Template" width="210"></a>

<ul> <li><code>Prompt Template</code> generates final prompts from template text. It handles <code>//</code> / <code>/*...*/</code> comment removal, wildcard expansion, Dynamic Prompts expansion, and <code>$key</code> replacement in one node.</li> <li>Wildcards are loaded from <code>ComfyUI/user/info_prompt_toolkit/wildcards</code> <code>.txt</code> files and support <code>__name__</code> (random) and <code>__name__#N</code> (fixed selection), including weighted lines in <code>weight::value</code> format.</li> <li>Dynamic Prompts supports <code>{a|b}</code> plus weighted options, multi-select, range selection, and the <code>@</code> cycle sampler.</li> <li>When <code>seed</code> is connected, random expansion becomes reproducible; without it, results are non-deterministic.</li> <li>When <code>extras</code> is connected, <code>$key</code> placeholders are replaced while missing keys are left as-is. <code>suffix</code> is appended after template expansion.</li> <li>Outputs include raw <code>string</code> and token-normalized <code>normalized_string</code>.</li> </ul> <br clear="left">

Image Reader / Image Saver

<a href="assets/readme/image-reader-and-image-saver.webp"><img align="left" hspace="16" src="assets/readme/image-reader-and-image-saver.webp" alt="Image Reader and Image Saver" width="210"></a>

<ul> <li><code>Image Reader</code> and <code>Image Saver</code> are core nodes that support reusable workflows for images that retain generation data (<code>A1111 infotext</code> / <code>image_info</code>).</li> <li><code>Image Saver</code> can save WebP or PNG images and embeds <code>A1111 infotext</code> into matching image metadata so prompt, model, and multiple LoRA references remain reusable.</li> <li><code>Check Referenced Models...</code> in the <code>Image Reader</code> right-click menu parses referenced entries from image infotext and lists Model / Refiner / Detailer / CLIP / VAE / LoRA items.</li> <li>The same window also shows local availability for each reference (Present / Missing), and can jump to <code>View Model Info...</code> for deeper inspection.</li> <li><code>Image Saver</code> supports both automatic serial naming and explicit <code>file_stem</code> naming, so you can switch naming strategy by workflow.</li> <li>When <code>write_caption</code> is enabled, a same-name <code>.txt</code> caption file is written alongside each image.</li> </ul> <br clear="left">

Referenced Image Saver

<a href="assets/readme/referenced-image-saver.webp"><img align="left" hspace="16" src="assets/readme/referenced-image-saver.webp" alt="Referenced Image Saver" width="210"></a>

<ul> <li><code>Referenced Image Saver</code> copies an existing file supplied through <code>image_reference</code> into the ComfyUI output directory and rewrites its A1111 infotext metadata from <code>image_info</code> without re-encoding the image pixels.</li> <li>It supports PNG / WebP / JPEG metadata updates and preserves the source extension, making it useful for saving lossy WebP or JPEG images again without further quality loss.</li> <li>When <code>file_stem</code> is unconnected, it uses the same automatic serial naming as <code>Image Saver</code>. When connected, it saves as <code>&lt;file_stem&gt;&lt;source extension&gt;</code> and overwrites an existing file with the same name.</li> <li><code>output_subdir</code> can organize files by date, and enabling <code>write_caption</code> writes a same-name <code>.txt</code> caption alongside each image.</li> </ul> <br clear="left">

Image Directory Reader

<a href="assets/readme/image-directory-reader.webp"><img align="left" hspace="16" src="assets/readme/image-directory-reader.webp" alt="Image Directory Reader" width="210"></a>

<ul> <li><code>Image Directory Reader</code> loads multiple images from subdirectories under ComfyUI <code>input</code> / <code>output</code> and returns corresponding data as aligned list outputs.</li> <li>You can choose the source with <code>path_source</code> and <code>path</code> (relative directory), then control the selection range with <code>start_index</code> and <code>image_load_limit</code> (<code>0</code> = unlimited).</li> <li>Target images are selected after case-insensitive filename sorting, making batch selection results easier to reproduce.</li> <li>For each image, it outputs <code>image</code>, <code>image_info</code> reconstructed and normalized from A1111 infotext, extensionless filename <code>file_stem</code>, and same-name <code>.txt</code> <code>caption</code> (empty string when missing).</li> <li>It also outputs <code>path_source</code> and <code>path</code>, so it can be connected directly to <code>Caption File Saver</code>.</li> </ul> <br clear="left">

Video Reader / Video Saver

<a href="assets/readme/video-reader-and-video-saver.webp"><img align="left" hspace="16" src="assets/readme/video-reader-and-video-saver.webp" alt="Video Reader and Video Saver" width="210"></a>

<ul> <li><code>Video Reader</code> and <code>Video Saver</code> are a paired set of nodes that support video reuse workflows while retaining generation metadata (<code>source_image_info</code> / <code>video_info</code>).</li> <li><code>Video Reader</code> outputs <code>image</code> as a list. If downstream nodes expect batch-oriented image input or do not assume list expansion, route through <code>Image List To Batch</code>.</li> <li>In <code>Video Saver</code>, the active <code>crf</code> field controls the quality/file-size tradeoff; lower values generally produce higher quality and larger files.</li> <li><code>source_image_info</code> keeps source-image information, while <code>video_info</code> keeps video-related information for easier review and comparison later.</li> <li><code>Video Reader</code> / <code>Video Saver</code> are available in environments where <code>PyAV</code> (<code>av</code>) is installed. Available codecs depend on the user's PyAV / FFmpeg build.</li> </ul> <br clear="left">

Mask Overlay Comparer

<a href="assets/readme/mask-overlay-comparer.webp"><img align="left" hspace="16" src="assets/readme/mask-overlay-comparer.webp" alt="Mask Overlay Comparer" width="210"></a>

<ul> <li><code>Mask Overlay Comparer</code> displays a slider-based comparison between the original <code>image</code> and a mask-overlayed view, making mask coverage easier to inspect visually.</li> <li>The comparer switches between two views: an easier-to-read background image and the image with mask overlay applied.</li> <li>The <code>mask</code> is visualized without binarization, using raw values in the <code>0.0..1.0</code> range.</li> <li>Visualization behavior is: <code>mask = 0.0</code> keeps the image unchanged, <code>mask = 1.0</code> renders black, and intermediate values are shown as a blue-tinted overlay.</li> <li>For preview generation, use <code>image</code> and <code>mask</code> at the same resolution.</li> </ul> <br clear="left">

Aspect Ratio to Size

<a href="assets/readme/aspect-ratio-to-size.webp"><img align="left" hspace="16" src="assets/readme/aspect-ratio-to-size.webp" alt="Aspect Ratio to Size" width="210"></a>

<ul> <li><code>Aspect Ratio to Size</code> recalculates <code>width</code> / <code>height</code> from node-UI widget edits before execution, so you can adjust values while checking <code>actual_ratio</code>.</li> <li>It resolves size from <code>width_ratio : height_ratio</code> and a base size, using either <code>width</code> or <code>height</code> as the anchor and deriving the other side from the ratio.</li> <li>The resolved size is aligned to <code>min_unit</code> (8 / 16 / 32 / 64), which helps keep resolutions on practical step sizes.</li> <li><code>width x height</code> is a combined size output that carries width and height on a single line, making it easy to pass into nodes that accept size input.</li> </ul> <br clear="left">

Aspect Ratio to Size (Trim Margin) / Trim Image by Margin

<a href="assets/readme/trim-margin-nodes.webp"><img align="left" hspace="16" src="assets/readme/trim-margin-nodes.webp" alt="Aspect Ratio to Size (Trim Margin) and Trim Image by Margin" width="210"></a>

<ul> <li><code>Aspect Ratio to Size (Trim Margin)</code> calculates a sampling size that extends beyond the desired image edges so that <code>Trim Image by Margin</code> can remove the margins afterward. The horizontal and vertical margin pairs stay linked, with each pair totaling either zero or <code>min_unit</code>. Its size outputs include all four margins and remain divisible by <code>min_unit</code>.</li> <li>Edit <code>actual_width</code> / <code>actual_height</code> as the size remaining after trimming. The readonly <code>width</code> / <code>height</code> fields show the sampling size including margins, while <code>actual_ratio</code> describes the trimmed size.</li> <li>The <code>margin</code> output bundles <code>top</code>, <code>right</code>, <code>bottom</code>, and <code>left</code> for <code>Trim Image by Margin</code>. When <code>margin</code> is unconnected, the trim node passes the image through unchanged.</li> <li><code>Set Margin Extra</code> / <code>Get Margin Extra</code> store margins in image-info extras and restore both the bundled margin and its four edge values.</li> </ul> <br clear="left">

Load New Model / Use Loaded Model

<a href="assets/readme/load-new-model-and-use-loaded-model.webp"><img align="left" hspace="16" src="assets/readme/load-new-model-and-use-loaded-model.webp" alt="Load New Model and Use Loaded Model" width="210"></a>

<ul> <li><code>Load New Model</code> and <code>Use Loaded Model</code> are a paired workflow that separates model loading from model reuse.</li> <li><code>Load New Model</code> receives the selected model from selector nodes such as <code>Checkpoint Selector</code> or <code>Diffusion Model Selector</code>. Checkpoint selections can load <code>model</code>, <code>clip</code>, and <code>vae</code> from the checkpoint; diffusion model selections load the <code>model</code> and use the optional <code>clip</code>, <code>vae</code>, and <code>vae_fallback</code> inputs for the accompanying CLIP/VAE runtime.</li> <li>When the same CLIP configuration is used across diffusion models and the CLIP remains unmodified, <code>Load New Model</code> can reuse the CLIP retained by a <code>Use Loaded Model</code> bundle through a weak-reference index and avoid loading it again during model changes.</li> <li><code>Use Loaded Model</code> keeps the final runtime bundle in a process-local cache keyed by the selected model, runtime settings, <code>lora_stack</code>, <code>apply_lora_to_clip</code>, and connected <code>clip</code> / <code>vae</code> conditions. On cache hit, the lazy <code>loaded_model</code> / <code>loaded_clip</code> / <code>loaded_vae</code> branch is skipped.</li> <li><code>Use Loaded Model</code> applies <code>lora_stack</code> internally by default. Set <code>apply_lora_to_clip=false</code> to apply it only to the model and keep CLIP reusable. Set <code>apply_lora_stack=false</code> if <code>loaded_model</code> and <code>loaded_clip</code> already include the same external LoRA processing.</li> <li>When placing <code>Lora Stack Lorader</code> or <code>TorchCompile</code>-related nodes before <code>Use Loaded Model</code>, connect their final outputs to the lazy inputs and set <code>apply_lora_stack=false</code>. <code>Lora Stack Lorader</code> also has <code>apply_lora_to_clip</code>; for external processing, set the same option on <code>Use Loaded Model</code> to describe the final CLIP state.</li> <li>Cache behavior is configurable in <code>ComfyUI Settings &gt; Info-Prompt-Toolkit &gt; Use Loaded Model Cache</code>. <code>Auto</code> always keeps the latest entry and prunes older entries by estimated memory size and <code>Memory budget ratio</code>; <code>Fixed</code> keeps up to <code>Max cache entries</code> regardless of the memory budget. <code>Max cache entries</code> defaults to <code>1</code> and can be set from <code>1</code> to <code>9</code>.</li> <li><code>Use cache logging</code> prints cache hit / miss / store / evict / clear messages to the ComfyUI Python console, including estimated memory amounts for store and evict events.</li> <li><code>lora_stack</code> matching is order- and strength-sensitive, so changing order or values is treated as a different condition set.</li> </ul> <br clear="left">

Model Merge

<a href="assets/readme/model-merge.webp"><img align="left" hspace="16" src="assets/readme/model-merge.webp" alt="Model Merge" width="210"></a>

<ul> <li><code>Model Merge</code> records two <code>IPT-Model</code> selections and their merge ratios in a new <code>IPT-Model</code> value. It does not load or merge model weights immediately; the actual model runtimes are loaded and merged when the result is passed to <code>Load New Model</code>.</li> <li>Connect the result to the corresponding <code>model</code>, <code>refiner</code>, or <code>detailer</code> input of <code>Image Info Context</code> or another image-info node to retain the complete merge configuration in <code>image_info</code>, including both source model references, their model types, and the model / CLIP ratios.</li> <li>When that <code>image_info</code> is saved by <code>Image Saver</code> or another compatible saver, the configuration is serialized as JSON in the A1111 infotext metadata. Depending on its <code>image_info</code> field, it is written as <code>Model Merge</code>, <code>Refiner Merge</code>, or <code>Detailer Merge</code> instead of being flattened into an ordinary single-model entry.</li> <li>When the saved image is loaded by <code>Image Reader</code>, the merge metadata is parsed back into the corresponding <code>model</code>, <code>refiner</code>, or <code>detailer</code> value in <code>image_info</code>. <code>Image Info Context</code> can then output the restored <code>IPT-Model</code> value for reuse with <code>Load New Model</code>.</li> <li>The metadata stores the merge configuration, not the merged model weights. The referenced source model files must therefore be available locally when the configuration is loaded and executed again.</li> <li><code>base_model</code> and <code>merge_model</code> must both be checkpoints or both be diffusion models; mixing the two model types is not supported.</li> <li><code>model_ratio</code> controls the model merge from the base model at <code>0.0</code> to the second model at <code>1.0</code>. For checkpoints, <code>clip_ratio</code> independently controls the CLIP merge over the same range.</li> <li>Checkpoint merges retain the VAE from <code>base_model</code>. For diffusion models, <code>clip_ratio</code> is ignored and CLIP / VAE can be supplied to <code>Load New Model</code> as needed.</li> </ul> <br clear="left">

Release Memory

<a href="assets/readme/release-memory.webp"><img align="left" hspace="16" src="assets/readme/release-memory.webp" alt="Release Memory" width="210"></a>

<ul> <li><code>Release Memory</code> releases selected persistent runtime references and requests Python / CUDA cleanup after the step connected to <code>after</code> has completed.</li> <li>The <code>after</code> input and <code>then</code> output are pass-through sockets, so the node can be placed near the end of a workflow without changing the data being passed onward.</li> <li><code>Generation Runtime</code> unloads ComfyUI generation models where possible and clears this extension's <code>Use Loaded Model</code> cache and recently loaded VAE runtime.</li> <li><code>SAM3 Runtime</code> clears the SAM3 Prompt To Mask processor cache, while <code>Tagger Runtime</code> clears PixAI Tagger and OppaiOracle Tagger torch model caches.</li> <li><code>GC &amp; CUDA Cleanup</code> runs garbage collection and asks the available CUDA cleanup hooks to release allocator-held memory where possible.</li> <li>The <code>Release Now</code> button performs the same cleanup manually when no prompt is currently running.</li> </ul> <br clear="left">

Selector Nodes

<a href="assets/readme/selector-nodes.webp"><img align="left" hspace="16" src="assets/readme/selector-nodes.webp" alt="Selector nodes" width="210"></a>

<ul> <li>Selector nodes centralize choices for reusable workflow inputs such as checkpoints, diffusion models, LoRAs, CLIP models, VAEs, samplers, and schedulers.</li> <li>Model-related selectors include <code>Checkpoint Selector</code>, <code>Diffusion Model Selector</code>, <code>Unet Model Selector</code>, <code>Lora Selector</code>, <code>Power Lora Selector</code>, <code>Vae Selector</code>, <code>CLIP Selector</code>, <code>Dual CLIP Selector</code>, <code>Triple CLIP Selector</code>, and <code>Quadruple CLIP Selector</code>.</li> <li>Supported model-related selectors provide a right-click <code>View Model Info...</code> action that uses the selected asset's SHA256 as a key to fetch metadata from Civitai and display it in a model information window.</li> <li>For <code>Lora Selector</code>, <code>View Model Info...</code> can also show a LoRA tag list with tag frequencies when metadata has been analyzed for the selected LoRA.</li> <li><code>Power Lora Selector</code> builds a LoRA stack from any number of rows; each row provides ON/OFF, LoRA, strength, model information, and removal controls.</li> <li>Common connection targets include <code>Image Info Defaults</code>, <code>Image Info Context</code>, <code>Load New Model</code>, <code>Use Loaded Model</code>, and <code>XY Plot Modifier</code>.</li> </ul> <br clear="left">

Example Workflows

Save Images With Metadata Using Image Saver

<a href="example_workflows/Save%20Images%20With%20Metadata%20Using%20Image%20Saver.webp"><img src="example_workflows/Save%20Images%20With%20Metadata%20Using%20Image%20Saver.webp" alt="Workflow sample: save images with metadata using Image Saver" width="900"></a>

Use Image Saver at the end of a generation workflow to save images with reusable metadata. The embedded A1111 infotext / image_info keeps prompt, model, LoRA, sampler, and related settings available for later reuse.

Generate From Metadata Loaded by Image Reader

<a href="example_workflows/Generate%20From%20Metadata%20Loaded%20by%20Image%20Reader.webp"><img src="example_workflows/Generate%20From%20Metadata%20Loaded%20by%20Image%20Reader.webp" alt="Workflow sample: generate from metadata loaded by Image Reader" width="900"></a>

Load a metadata-rich image with Image Reader, pass the restored image_info into the workflow, and generate from the saved settings as a starting point for reproduction or further iteration.

Filter and Encrypt Metadata Before Saving With Referenced Image Saver

<a href="example_workflows/Filter%20and%20Encrypt%20Metadata%20Before%20Saving%20With%20Referenced%20Image%20Saver.webp"><img src="example_workflows/Filter%20and%20Encrypt%20Metadata%20Before%20Saving%20With%20Referenced%20Image%20Saver.webp" alt="Workflow sample: filter and encrypt metadata before saving with Referenced Image Saver" width="900"></a>

Load an existing image with Image Reader, remove selected main fields and extra keys from its image_info, specify which remaining fields to encrypt, and save the result with Referenced Image Saver. This lets you choose which metadata is removed, kept public, or stored in encrypted form without re-encoding the image pixels.

License

• The original source code of this project is offered under GPL-3.0-or-later.
• This repository also includes vendored sam3, which is distributed under the SAM License.
• See THIRD_PARTY_LICENSES.md for third-party license details, including notes on separately obtained optional models.
• Model files and other external assets obtained separately by users remain subject to their respective upstream licenses and terms.