cgem156-ComfyUI🍌

Regional prompting, the original hard-mode version

An experimental attention-temperature knob, filed as a test node

Block specific tags from the Dart tag generator

The manual entry point into cgem156's batch-prompt pipeline

Push detail and contrast without a LoRA or a finetune

Encode a whole batch of prompts in one node

Split what the model attends to from what it renders

The generation knobs for the tag-completion LLM

Turn a handful of tags into a full Danbooru-style prompt

Build the structured seed string Dart actually expects

The structured seed builder with aspect ratio and identity control

Heatmap every tag above a confidence threshold, no target needed

See exactly where the tagger is looking when it assigns a tag

A heatmap and its tag, matched up automatically

A hi-res fix baked into the sampling loop instead of bolted on after

Smarter color reduction than a flat posterize

The stock advanced sampler, wired to accept a whole batch of models

The other half of the LoRA-strength sweep

Pushing LCM past its usual CFG 1-2 ceiling without the usual burn

Only pay for guidance during the steps that actually need it

The loader for a tiny LLM that finishes your Danbooru tags for you

The loader for WD-tagger, the anime auto-captioning model

Apply a pre-defined LoRA to your model

Test a LoRA at ten strengths in one run

Define a LoRA's weight without applying it yet

Combine two LoRAs into one file, algebraically

The write-to-disk step for cgem156's LoRA merger pipeline

Shrink a LoRA's rank without hand-picking a number

The author's own words for it are 'a mysterious ControlNet'

See exactly where two latents actually differ

MSE Heatmap's cousin, working in the tagger's feature space

Ten LoRA slots in one node, with a normalize switch

Three LoRAs on one node, with a normalize switch that saves you from overcooking

Five LoRA slots on one node, and why you'll actually want the normalize switch this time

Reference-only attention injection, but from more than one reference at once

The batch-builder the Reference Apply nodes need to actually do anything

Run WD Tagger v3 on an image to get Danbooru tags

Add the same text to the front of every prompt in a batch

The closest thing ComfyUI has to A1111's built-in XYZ Plot

Batch-generate prompt variants with different colors

A seeded starting noise with one extra dial

A1111's variation-seed trick, ported to ComfyUI's noise pipeline

The old reference-only ControlNet trick, reimplemented as a model patch

Pull one item out of a batch and re-batch it

Keeping a batch's color and tone consistent, channel by channel

The manual sampler node, with a preview image built in

Running a whole parameter grid through one sampler node

Euler A's variety, without losing your ability to A/B test

Dump a batch of prompts to text files

Write a whole batch to disk with matching filenames

Skip the PNG, cache raw latents to disk instead

Generating past your model's native resolution without a second pass

Searching for a noise draw that stays close to a reference

The smallest possible node, and why that's useful

The Hyper-SD sampler, on its own, with one dial

Type your own sigma schedule instead of picking one from a dropdown

The most undocumented corner of an already terse pack

A1111's variation seed and strength, in full, for ComfyUI's noise pipeline

Find related Danbooru tags, not tag an image