ComfyUI_EmAySee_CustomNodes

Five prompt skeletons, one seed — EmAySee's fill-in-the-blank prompt shaker

Point ComfyUI at a local LLM — the kitchen-sink Oobabooga chat connector

The Oobabooga connector that explains itself — tooltips on every knob

A wire that got a name — when you just need to route anything through ComfyUI

You know the box, skip the detector — hand-drawn bounding boxes into SEGS

Seventeen anatomy sliders in one box — organized floats for character morphing

The text twin — seventeen attribute strings on one node for prompt building

The mid-workflow 'did you try turning it off and on' — forcing a VRAM purge

One number in, two out — the aspect-ratio calculator that keeps things proportional

Checkbox to Float — a boolean in, a 1.0 or 0.0 out, zero surprises

Only resize the small ones — the upscale gate that leaves big images alone

The original context loop — rolling LLM memory written to a plain text file

The chatty rewrite — Context Manager V2 and its console-debug spam

Context Reader V5 — the read half of ComfyUI's file-based LLM memory

Context Writer V5 — the append half that fixes your LLM's headers for you

The watchdog writer — absolute paths, bounded history, and a truncation tripwire

A timestamp that double-duties as a filename — and a seed input that does nothing

An LLM connector that hands you the thinking and the answer separately — built for local oobabooga-style backends

The same reasoning LLM connector, now with a seed you can actually fix

Flip width and height without rewiring your graph

Inline wildcard syntax with a seed — dynamic prompts that you can actually reproduce

Drive any slider from a 0–1 value — percent in, real range out

One slider, many string inputs — pick which prompt wins without rewiring

The blank canvas for Qwen-Image workflows — layers included, reference sampling left out

A preview node that won't fill your temp folder — for long monitoring runs

Turn coded filenames into prompt triggers — 'img_02_v1.png' becomes 'young adult'

A float, dressed up as prompt text — strength values without the typing

Ask a loaded model where it came from — handy for logs, fragile under the hood

A text file on disk, disguised as a global variable — share state between workflows

Save workflow state to disk — with a built-in append-only history

A > B? This node gives you a 1 or a 0 to branch on

Is your API up? This node pings it and hands you a 1 or 0

Give me the width and height — the one-liner image dimension node

A run log that appends itself — prompt, filename, and state, every generation

A zero-work node that does nothing — and that's exactly its job

Preview the image without killing the wire — see it and keep it flowing

It doesn't pad the image — it tells you how much to pad, and that's more useful

See the caption next to the image, and pass both through untouched

Pick one of four strings with a slider — your prompt-switcher

Twenty prompt slots under one selector — the big sibling

The latent switch that isn't a switch — it's a priority override

A vision LLM node that isn't Llama and doesn't need a key — point it at Ollama

Strip the <think> blocks out of your LLM's chat

Load an image and get its filename out — the dataset workhorse

Apply a LoRA whose name comes in as plain text

Pick a LoRA, get its filename AND its absolute path

The loader half of the two-piece LoRA setup

Read a LoRA's training secrets straight off the safetensors header

The same LoRA metadata dump, minus the tag-frequency wall of text

Turn a dropdown choice into a ready-to-apply LORA_OBJECT

Nine LoRAs pre-loaded, one selector decides which applies

A LoRA dropdown and two strength knobs, all as outputs

Stack up to twenty LoRAs, applied in a fixed order

Punch a rectangle out of the middle of your subject

Grow or shrink a mask — with independent X and Y

Grow a mask, X and Y separately — expand-only edition

A mask that refuses to fail — always hands you an image

See what your mask actually covers before you send it to a sampler

Shrink a mask on its narrow axis — a niche tool with a clear job

Merge up to 14 masks into one — the union node that just works

Expand, shrink, and punch holes in one node

Max(w, h) and friends — the mini calculator that speaks Python

Scale to a target megapixel count, not a fixed resolution

Turn your LoRA's training JSON into something a human can read

Ask ComfyUI what models are actually sitting in VRAM right now

Bundle 20 body-morph sliders into one tidy pipe

Unpack a morph pipe back into its twenty sliders

Ten detectors, one SEGS — running every bbox model at once in ComfyUI

Get one bounding box around all your masks at once

Two integers, one slider, kept in ratio — the tiny utility that earns its place

A nine-slot switch that hands you both the index and the label

Grow your canvas and get the mask that marks the new space

Pick a string at random, but make one option rare on purpose

' and 'STATE:' reply into two clean strings

Resize to a Qwen-friendly resolution before the model sees your image

Caption any image into a rich prompt — locally with Qwen-VL or via any OpenAI-compatible API

Pin an image to Qwen's native aspect ratio before it goes downstream

Snap your image to the nearest Qwen aspect ratio, scaled to taste

The fixed Qwen resolution optimizer — use this one, not the original

That zoom you hate on Qwen Image Edit is a resolution problem

A local vision LLM that turns your images into prompts — if you've got the VRAM

The dumbest possible random integer node — and sometimes that's exactly right

Pick from only the options you switched on

The random integer node that actually understands ranges

Roll a random one from your pool

Portrait or landscape? Let the seed decide

Twenty prompt slots, one randomly chosen — the 20-way picker

A coin flip between two prompts, wrapped in a node

Four prompts in, one random out — the 4-way picker

Random from three strings, because sometimes a coin isn't enough

Dedupe a comma-separated string without touching a spreadsheet

Dedupe a word list — and accept that it's going to sort itself

The dedup node that actually respects your order

Surgically remove words from a prompt — whole words, not substrings

A KSampler that repaints only what your mask says — no extra nodes

The Save Image that's allowed to leave ComfyUI's output folder

Save to an exact path with an exact filename — built for dataset dumps

Get that LLM output off the graph and onto disk as a .txt

It doesn't make audio — it asks your local TTS server to

Merge every SEGS stream in your graph into one — without re-running the detector

Push a CPU-offloaded model back onto your GPU

Free one model's VRAM without nuking the checkpoint you still need

Send an image to a local vision LLM and get a caption back as a prompt

Does this string have text in it?

A five-way string switch for picking the pose descriptor mid-workflow

The one node the README itself tells you not to use

Call your local text-generation-webui from inside a workflow — no API key needed

The same local-LLM call, but it sends a Bearer token your server asked for

The `_test` variant — same node, extra knob, author's own servers in the defaults

Chat API call that splits the model's reasoning out of its answer

The kitchen-sink chat API node with thinking-tag parsing — the one you'd actually use

Split an LLM's [NatDes]/[TagDes] output into description and tags

Strip the tags you don't want out of a booru-style prompt, case-insensitively

Join up to 50 strings with newlines before they hit your prompt encoder

The Qwen-Image-Edit text encoder, minus the resolution-selector wiring

Up to seven references, one instruction

Twenty find-and-replace pairs in one node, applied in order

One checkbox, two integers — a boolean-to-int switch for conditional values

Twelve body-attribute sliders that turn into prompt text lines

The body-attribute text builder

The body-attribute slider node that hands you every string separately

Match a reference image's aspect ratio without tripping the VAE

%var% placeholders

The 'REALLY Unique String Selector' is just a two-way string switch — and that's fine

Save an image with its prompt in a sidecar .txt — know the overwrite gotcha

Strip the boilerplate — cleaning WD14 tagger output before it hits your prompt