| Pack | Nodes | |------|-------| | comfyui-generator-pack | Tags suggested from a Danbooru dataset, sentence-to-tags and character tags: Tags Generator, Tags Extractor, Character Tags Generator, Tags Conflict Filter, Classify Tags, Group Tags | | comfyui-daam-pack | Cross-attention heatmaps per prompt tag: Sampler Custom (DAAM), DAAM Tag Explorer | | comfyui-danbooru-pack | Danbooru tag retrievers and posts downloader | | comfyui-evaluate-pack | Evaluate: transforms a string with Python code written in the node | | comfyui-api-pack | External APIs: zero-shot classification, OpenAI Inference, remote ComfyUI API (Api Generate / Submit / Collect, Load Workflow), Grok image-to-video | | comfyui-filter-pack | Image filters: Adjust Image (colour correction, sharpen, denoise, resize) |

Nodes that moved out of this pack kept their node ids, so existing workflows load once the pack that owns them is installed.

Installation

Clone or copy this repository into the custom_nodes directory of your ComfyUI installation.
Restart ComfyUI.

Provided Nodes

Prompt Nodes (`AlcheminePack/Prompt`)

Prompt Workflow

| Node | Description | |------|-------------| | ProcessTags | Full pipeline for tag processing. Combines ReplaceUnderscores → FilterTags → FilterSubtags → FilterColors → FilterPlurals → SDXLAutoBreak in sequence. | | FilterTags | Removes blacklisted tags from prompts. Supports wildcards defined in resources/wildcards.yaml. | | FilterSubtags | Removes duplicate/unnecessary subtags (e.g., dog, white dog → white dog). | | ReplaceUnderscores | Converts all underscores (_) to spaces. | | FixBreakAfterTIPO | Fixes BREAK token formatting after TIPO output (removes weights like (BREAK:-1)). | | SDXLTokenAnalyzer | Analyzes CLIP tokens in a prompt (SDXL only). Returns g/l tokenizer results with token counts. | | RemoveWeights | Removes all weight notations from tags (e.g., (cat:1.2) → cat). | | FilterColors | Keeps one colour per thing: of red dress, blue dress, the first. Colours come from the color key of resources/wildcards.yaml. | | FilterPlurals | Of two tags that differ only by a plural s (arm up, arms up), keeps the first. | | BoySubjectFilter | When a tag needs a man (sex, hetero, penis, anything spelled with another), removes solo and, if no boy is counted, adds ((1boy)) and add_tags. | | SDXLAutoBreak | Automatically inserts BREAK to keep each segment within 75 tokens (SDXL only). | | SubstituteTags | Regex-based tag substitution with conditional execution (run_if, skip_if). | | SeparateLoraTags | Separates lora tags (<lora:...>) from a prompt. If the same lora appears multiple times, the last weight is used. | | TextPrompt | Plain multiline text input with dynamicPrompts off, so typing {a|b} no longer jumps the cursor to the end. Wildcards are expanded in Python instead. |

ProcessTags

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | text | STRING | (required) | Input prompt text | | replace_underscores | BOOLEAN | True | Replace underscores with spaces | | filter_tags | BOOLEAN | True | Remove blacklisted tags | | filter_subtags | BOOLEAN | True | Remove duplicate/unnecessary subtags | | filter_colors | BOOLEAN | True | Keep one colour per thing (FilterColors) | | filter_plurals | BOOLEAN | True | Keep the first of two tags that differ only by a plural s (FilterPlurals) | | auto_break | BOOLEAN | False | Auto-insert BREAK for 75-token limit | | clip | CLIP | (optional) | Required for auto_break | | blacklist_tags | STRING | "" | Comma-separated blacklist (supports wildcards) | | fixed_tags | STRING | "" | Tags to preserve regardless of filtering |

| Output | Description | |--------|-------------| | processed_text | The processed prompt text | | filtered_tags_list | List of removed-tag groups (one entry each from the FilterTags / FilterSubtags / FilterColors / FilterPlurals steps) |

⚠️ 6.0.0: filter_colors and filter_plurals sit between filter_subtags and auto_break. In a workflow saved before 6.0.0 the values of auto_break, blacklist_tags and fixed_tags shift by two widgets; set them again. An API-format workflow needs the two new inputs added.

FilterTags

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | text | STRING | (required) | Input prompt text | | blacklist_tags | STRING | "" | Comma-separated blacklist (supports wildcards) | | fixed_tags | STRING | "" | Tags to preserve regardless of filtering |

| Output | Description | |--------|-------------| | processed_text | Prompt with blacklisted tags removed | | filtered_tags | Comma-separated list of the removed tags |

FilterSubtags

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | text | STRING | (required) | Input prompt text | | fixed_tags | STRING | "" | Tags to preserve regardless of subtag filtering |

| Output | Description | |--------|-------------| | processed_text | Prompt with redundant subtags removed | | filtered_tags | Comma-separated list of the removed subtags |

SDXLTokenAnalyzer

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | clip | CLIP | (required) | CLIP model (must expose clip_g / clip_l) | | text | STRING | (required) | Input prompt text |

| Output | Description | |--------|-------------| | g_tokens | Tokens decoded by clip_g (segments separated by BREAK) | | g_token_count | Per-segment clip_g token counts, comma-separated | | l_tokens | Tokens decoded by clip_l | | l_token_count | Per-segment clip_l token counts, comma-separated |

SDXLAutoBreak

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | clip | CLIP | (required) | CLIP model (uses clip_g token count) | | text | STRING | (required) | Input prompt text |

ReplaceUnderscores / FixBreakAfterTIPO / RemoveWeights

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | text | STRING | (required) | Input prompt text |

FilterColors / FilterPlurals

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | text | STRING | (required) | Input prompt text |

| Output | Description | |--------|-------------| | processed_text | Prompt with the repeated tags removed | | filtered_tags | The removed tags |

BoySubjectFilter

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | text | STRING | (required) | Input prompt text | | add_tags | STRING | "(hetero:1.1), (couple:1.1), (deep skin:1.1)" | Tags added after ((1boy)) when no boy is counted |

A prompt without a tag that needs a man is returned unchanged. Tags are matched whole (sex toy, sexy do not match), tags starting with after are skipped, and underscores and case are ignored.

SubstituteTags

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | text | STRING | (required) | Input prompt text | | pattern | STRING | "" | Regex pattern to match | | repl | STRING | "" | Replacement string | | run_if | STRING | "" | Only run if this pattern exists | | skip_if | STRING | "" | Skip if this pattern exists |

SeparateLoraTags

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | text | STRING | (required) | Input prompt text |

| Output | Description | |--------|-------------| | text_without_lora | Text with all lora tags removed (whitespace/newlines preserved as much as possible) | | text_with_lora | Deduplicated lora tags joined by spaces (when the same lora appears multiple times, the last weight wins) |

TextPrompt

ComfyUI's built-in text widget has dynamicPrompts enabled, which re-parses the field as you type: writing {a|b} moves the cursor to the end on every keystroke. This node turns the flag off, so the widget behaves like a plain text box, and resolves the {option1|option2|...} syntax in Python at execution time instead (one random pick per group, nesting supported).

When ComfyUI-Impact-Pack is installed, its wildcard engine is used instead, so __wildcard__ file references, $$ multi-select and # comments are also resolved — a TextPrompt can stand in for an ImpactWildcardProcessor node. Without it, only the {a|b} syntax is resolved.

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | text | STRING | (required) | Multiline prompt text; {a|b} groups are expanded on execution | | seed | INT | 0 | Input-only; seeds the wildcard picks so a run is reproducible |

| Output | Description | |--------|-------------| | text | Prompt with every wildcard group resolved |

Everywhere Nodes (`AlcheminePack/Everywhere`)

| Node | Description | |------|-------------| | Everywhere | A cg-use-everywhere broadcaster with fixed, named inputs (model, clip, vae, positive, negative, latent_image, seed, blacklist). The pinned input names keep UE's name-based routing working, so the two conditionings stay apart without renaming slots by hand. Requires cg-use-everywhere. |

Flow Control Nodes (`AlcheminePack/FlowControl`)

Flow Control Workflow

| Node | Description | |------|-------------| | Lazy Execution | Passes value through only after signal resolves. Controls execution order, and propagates an upstream ExecutionBlocker on signal to gate downstream nodes. |

Lazy Execution

| Parameter | Type | Description | |-----------|------|-------------| | value | ANY | Value to pass through | | signal | ANY | Gate input — value is only forwarded once this resolves |

| Output | Description | |--------|-------------| | value | The value input, forwarded once signal has resolved |

Use Case: When you need sequential execution (e.g., run generation B only after generation A completes), or want a downstream branch to be skipped while an upstream node returns an ExecutionBlocker (e.g. gate Api Submit from comfyui-api-pack on Api Collect so a new job is only submitted once the previous one finishes).

Muting to prime a gated graph: because value is the first input, muting (bypassing) this node passes value straight through, ignoring signal. This is handy for cold-starting a loop that would otherwise be blocked forever — e.g. when Api Collect has no job yet and keeps emitting an ExecutionBlocker, mute the gate for one run to fire the first Api Submit, then un-mute it to resume normal gating.

Lora Nodes (`AlcheminePack/Lora`) (Experimental)

| Node | Description | |------|-------------| | DownloadImage | Downloads an image from a URL into the ComfyUI output directory. | | LoadImagesFromDir | Loads every image of a folder under the ComfyUI output directory as a list, with the file names, so a dataset folder can be addressed by the same relative path as SaveImageWithText and the downloader nodes. Re-runs when the folder's files change. | | SaveImageWithText | Saves an image alongside a .txt caption file (for LoRA training dataset preparation). Output node: runs even when nothing is connected downstream. With name it keeps the source file name and skips pairs that already exist. | | FilterFrequentTags | Drops the tags that appear in at least threshold of all captions, so the subject's constant traits are learned into the trigger word instead of the captions. Takes the whole caption list at once. | | TrainSdxlLora | Trains an SDXL LoRA with kohya sd-scripts installed next to ComfyUI (<ComfyUI>/sd-scripts with its own .venv). Everything lives under output/train/<name>/<version>/: train/ images and .txt captions are read in place, mask/ (if present) turns on masked loss, each training gets a timestamp folder under both checkpoints/ (LoRA files) and runs/. |

DownloadImage

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | url | STRING | (required) | Image URL to download | | dir_path | STRING | "output/images" | Destination directory (relative to ComfyUI output) |

| Output | Description | |--------|-------------| | image | Loaded image tensor | | file_path | Path of the saved file (relative to output dir) |

LoadImagesFromDir

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | dir_path | STRING | "" | Folder relative to the ComfyUI output directory |

| Output | Description | |--------|-------------| | images | One RGB image per file (list), file names in ascending order; jpg / jpeg / png / webp | | names | The file names, same order (feed SaveImageWithText.name to keep them) |

SaveImageWithText

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | image | IMAGE | (required) | Image to save | | text | STRING | (required) | Caption text saved as a sibling .txt | | dir_path | STRING | (required) | Destination directory (relative to ComfyUI output) | | prefix | STRING | "" | Filename prefix; auto-increments index when set | | name | STRING | (optional) | Source file name or path; its stem names the saved pair (raw/a.jpg → a.png + a.txt) instead of a running number | | overwrite | BOOLEAN | False | With name: rewrite an existing pair. Off, an existing pair is kept and its paths returned |

| Output | Description | |--------|-------------| | image_path | Path of the saved .png | | text_path | Path of the saved .txt |

FilterFrequentTags

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | captions | STRING (list) | (required) | All captions of the dataset, one per image | | threshold | FLOAT | 0.8 | A tag found in this share of the captions or more is removed | | fixed_tags | STRING | "" | Comma-separated tags never removed (the trigger word) |

| Output | Description | |--------|-------------| | captions | The captions with the frequent tags removed, same order | | removed_tags | One line like blue eyes(28/34), blonde hair(32/34) |

TrainSdxlLora

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | name | STRING | "my_lora" | Dataset and LoRA file name; the dataset is output/train/<name>/<version>/train/ (images with same-name .txt captions) | | version | STRING | "v1.0.0" | Dataset version folder under output/train/<name>/ | | ckpt_name | COMBO | (required) | SDXL checkpoint from models/checkpoints | | train_unit / train_length | COMBO / INT | epochs / 20 | How long to train: epochs (steps = epochs × ceil(images / batch size)) or steps as given | | learning_rate | FLOAT | 1.0 | UNet learning rate; a multiplier for Prodigy (keep 1.0), about 1e-4 for the AdamW optimizers | | network_type | COMBO | LoRA | LoRA or LoKr (Kronecker product, sd-scripts networks.lokr); rank, alpha, LoRA+ and dropout apply to both | | r / lora_alpha | INT | 32 / 16 | LoRA rank and alpha | | loraplus_lr_ratio | FLOAT | 0 | LoRA+: learning rate of the up matrices as a multiple of the down matrices (4-16 in the paper), 0 to disable | | resolution | INT | 1024 | Bucket resolution | | batch_size | INT | 1 | Batch size | | optim | COMBO | Prodigy | Prodigy (adaptive step size, safeguard_warmup, use_bias_correction, weight decay 0.01, betas 0.9/0.99), AdamWScheduleFree, AdamW8bit, AdamW | | lr_scheduler_type | COMBO | cosine | cosine or constant | | warmup_ratio | FLOAT | 0.05 | Warmup steps as a share of the total steps | | shuffle_caption / keep_tokens | BOOLEAN / INT | True / 1 | Shuffle tags each step, keeping the first keep_tokens in place | | gradient_checkpointing | BOOLEAN | False | Lower VRAM, slower | | gradient_accumulation_steps | INT | 1 | Optimizer steps are taken every N batches (effective batch = N × batch size) | | network_dropout | FLOAT | 0.0 | LoRA neuron dropout against overfitting, 0 to disable | | scale_weight_norms | FLOAT | 0.0 | Scale LoRA weights back when their norm exceeds this, 0 to disable | | noise_offset / noise_offset_value | BOOLEAN / FLOAT | True / 0.0357 | Noise offset (drawn as one row: toggle and value). 0.0357 is what the SDXL base was trained with; switch it off for AnyIllustrious-XL | | seed | INT | 42 | Training seed | | save_steps | INT | 200 | Intermediate LoRA files, 0 to disable | | sample_steps | INT | 200 | Sample images, 0 to disable | | logging_steps | INT | 25 | Send a point to the loss chart every N steps (the last step is always sent) | | sample_prompts | STRING (multiline) | 1girl, solo --n worst quality, low quality --w 832 --h 1216 --s 28 --l 5 --d 42 | One sd-scripts sample prompt per line; # lines are comments |

| Output | Description | |--------|-------------| | lora_path | The final .safetensors, output/train/<name>/<version>/checkpoints/<timestamp>/<name>.safetensors (snapshots from save_steps sit next to it) | | run_dir | output/train/<name>/<version>/runs/<timestamp>/: TOML configs, sample/ (step<step>_<timestamp>_<prompt line>.png), logs/ |

Always set in the training config: min_snr_gamma = 5, mixed_precision = "bf16", cache_latents, sdpa, UNet-only training.

Layout under output/train/<name>/<version>/:

| Folder | Role | |--------|------| | train/ | Images with same-name .txt captions, read in place | | mask/ | Optional same-name mask images (white = train, black = ignore); when the folder exists, sd-scripts masked_loss weights the loss by the mask's R channel | | checkpoints/ | One <timestamp>/ folder per training with its LoRA files | | runs/ | The same <timestamp>/ per training: TOML configs, sample/, logs/ |

Above the loss chart the node shows the output_dir and mask_dir it will use, updated as you type, and, once a training starts, the model_dir (checkpoints/<timestamp>) and the step count computed from the image count. While training runs, the node draws a live avr_loss chart (step and loss from the sd-scripts log).

Always on: bf16, sdpa, cached latents, UNet only, buckets, TensorBoard logging (<sd-scripts>/.venv/bin/tensorboard --logdir output/train/<name>/<version>/runs). The node waits for sd-scripts and streams its log to the console with a [sd-scripts] prefix.

Model Nodes (`AlcheminePack/Model`)

| Node | Description | |------|-------------| | Cached Load LoRA Tag | Loads the LoRAs referenced by <lora:name:weight> tags in the text and caches the patched result. |

Cached Load LoRA Tag

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | model | MODEL | (required) | Base model to patch | | clip | CLIP | (required) | Base CLIP to patch | | text | STRING (multiline) | (required) | Prompt containing <lora:...> tags |

| Output | Description | |--------|-------------| | MODEL | Model patched with the referenced LoRAs | | CLIP | CLIP patched with the referenced LoRAs | | STRING | The prompt with all lora tags stripped out |

Tag format: <lora:name:model_weight:clip_weight> — the clip weight is optional and defaults to the model weight. A tag without any numeric weight (<lora:name>) is not recognized as a lora tag: it is neither loaded nor stripped from the output text. Use <lora:name:0> to load at weight 0.
name is matched as a prefix against files in the loras folder; unmatched tags are skipped.
The patched result is cached while text/model/clip are unchanged, skipping LoRA re-loading and re-patching.

Wildcard Support

The FilterTags and ProcessTags nodes support wildcards defined in resources/wildcards.yaml.

Example: Using <color> in the blacklist expands to every color defined in the YAML file, joined into one alternation — (shiny|dark|colored|<color>) skin blocks blue skin, grey skin, two-tone skin and the rest in one line.

Angle brackets, not the __color__ form the wildcard packs use. A wildcard processor (ImpactWildcardProcessor, Dynamic Prompts) runs upstream of this node, so it resolves __color__ before FilterTags ever sees it — and it resolves it by picking one value, which is the opposite of what a blacklist wants. No wildcard processor claims <...>, so the token survives them and arrives here intact. The __color__ form still works for blacklists that never pass through one.

Examples

ProcessTags Example

Input: dog, cat, white dog, black cat
Blacklist: ^cat$
Output: white dog, black cat
Filtered: ['cat', 'dog']   (one entry per step: FilterTags, then FilterSubtags)

Blacklist tokens are regexes matched anywhere in a tag, so a bare cat would also remove black cat — anchor with ^cat$ to match the tag exactly.

FilterSubtags Example

Input: dog, cat, white dog, black cat
Output: white dog, black cat
(Removes 'dog' and 'cat' as they are subtags of 'white dog' and 'black cat')

SeparateLoraTags Example

Input:
moriaruruka, <lora:characters\lulurka\1-moriaruruka.safetensors:0.7> blonde, gradient hair, jewelry,
<lora:characters\lulurka\2-moriaruruka.safetensors:0.7> <lora:characters\lulurka\3-moriaruruka.safetensors:0.7> <lora:characters\lulurka\3-moriaruruka.safetensors:1.0>

Output:
text_without_lora: moriaruruka, blonde, gradient hair, jewelry
text_with_lora: <lora:characters\lulurka\1-moriaruruka.safetensors:0.7> <lora:characters\lulurka\2-moriaruruka.safetensors:0.7> <lora:characters\lulurka\3-moriaruruka.safetensors:1.0>

When a lora block is followed by ,, the preceding comma/whitespace is also removed to avoid double commas.
When a lora block is not followed by ,, the preceding comma is preserved and only the preceding whitespace is removed.
When the same lora appears multiple times, the last specified weight wins (e.g., the final weight for 3-moriaruruka.safetensors above is 1.0).

SubstituteTags Example

# If "girl" doesn't exist, replace "1boy" with "1girl, 1boy"
pattern: 1boy
repl: 1girl, 1boy
skip_if: girl

License

GPL-3.0 License

ComfyUI-Alchemine-Pack

Nodes (22)