# ComfyUI-Alchemine-Pack A custom node pack for [ComfyUI](https://github.com/comfyanonymous/ComfyUI) that provides utility nodes for prompt processing, LoRA-tag loading, workflow control, and input broadcasting. ## Related Packs | Pack | Nodes | |------|-------| | [comfyui-generator-pack](https://github.com/alchemine/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](https://github.com/alchemine/comfyui-daam-pack) | Cross-attention heatmaps per prompt tag: Sampler Custom (DAAM), DAAM Tag Explorer | | [comfyui-danbooru-pack](https://github.com/alchemine/comfyui-danbooru-pack) | Danbooru tag retrievers and posts downloader | | [comfyui-evaluate-pack](https://github.com/alchemine/comfyui-evaluate-pack) | Evaluate: transforms a string with Python code written in the node | | [comfyui-api-pack](https://github.com/alchemine/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](https://github.com/alchemine/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 1. Clone or copy this repository into the `custom_nodes` directory of your ComfyUI installation. 2. Restart ComfyUI. ## Provided Nodes ### Prompt Nodes (`AlcheminePack/Prompt`) ![Prompt Workflow](workflows/comfyui-alchemine-pack-workflow-Prompt.png) | 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 (``) 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](https://github.com/ltdrdata/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](https://github.com/chrisgoringe/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](workflows/comfyui-alchemine-pack-workflow-FlowControl.png) | 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](https://github.com/alchemine/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. --- ### Image Nodes (`AlcheminePack/Image`) | Node | Description | |------|-------------| | **FastImageSave** | Saves images as JPEG with the workflow embedded. Drop a saved file on the canvas to load its workflow. | #### FastImageSave | Parameter | Type | Default | Description | |-----------|------|---------|-------------| | `images` | IMAGE | (required) | Images to save, at quality 95 | | `filename_prefix` | STRING | `ComfyUI` | File name prefix under the output directory. Accepts subfolders and variables such as `%date:yyyy-MM-dd%/image` | | Output | Description | |--------|-------------| | `images` | The `images` input, forwarded unchanged | Only this pack reads the embedded workflow: other tools show it as an ordinary JPEG. --- ### 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 (`/sd-scripts` with its own `.venv`). Everything lives under `output/train///`: `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 | | `count` | INT | 0 | Images per run; 0 loads the whole folder | | `chunk` | INT | 0 | With `count`, which group of `count` images to load (0 = the first). Set it to increment and queue once per group | | 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///train/` (images with same-name `.txt` captions) | | `version` | STRING | "v1.0.0" | Dataset version folder under `output/train//` | | `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_unit` / `save_steps` | COMBO / INT | steps / 200 | Intermediate LoRA files every that many `epochs` or `steps`, 0 to disable | | `sample_unit` / `sample_steps` | COMBO / INT | steps / 200 | Sample images every that many `epochs` or `steps`, 0 to disable | | `logging_unit` / `logging_steps` | COMBO / INT | steps / 25 | Send a point to the loss chart every N `epochs` or `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///checkpoints//.safetensors` (snapshots from `save_steps` sit next to it) | | `run_dir` | `output/train///runs//`: TOML configs, `sample/` (`step__.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///`: | 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 `/` folder per training with its LoRA files | | `runs/` | The same `/` 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/`) 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 (`/.venv/bin/tensorboard --logdir output/train///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 `` 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 `` 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: `` — the clip weight is optional and defaults to the model weight. A tag without any numeric weight (``) is not recognized as a lora tag: it is neither loaded nor stripped from the output text. Use `` 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 `` in the blacklist expands to every color defined in the YAML file, joined into one alternation — `(shiny|dark|colored|) 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, blonde, gradient hair, jewelry,

Output: text_without_lora: moriaruruka, blonde, gradient hair, jewelry text_with_lora:

``` - 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

Fast Image Save

Inputs (2)

Outputs (1)