ComfyUI Bawk Nodes

Compact FLUX workflow nodes for ComfyUI

FLUX.1 dev · Krea · Kontext · FLUX.2 dev · FLUX.2 klein 4B / 9B · GGUF

A handful of nodes cover a full multi-LoRA FLUX workflow:

Bawk Model Loader ──────────────────┬─► Bawk Wildcard Encoder ─► Bawk Sampler ─► Bawk Image Saver
Bawk LoRA Stack (Set A) ─► Bawk LoRA Switch ─┘
Bawk LoRA Stack (Set B) ─┘

FLUX models do not use negative prompts, so there is none anywhere in this pack.

Nodes

🚀 Bawk Model Loader

Loads diffusion model, VAE and text encoder(s) in one node.

Model	clip_name1	clip_name2	vae_name
FLUX.1 dev / Krea / Kontext	t5xxl_*.safetensors	clip_l.safetensors	ae.safetensors
FLUX.2 dev	mistral_3_small_flux2_*.safetensors	none	flux2-vae.safetensors
FLUX.2 klein 4B	qwen_3_4b.safetensors	none	flux2-vae.safetensors
FLUX.2 klein 9B	qwen_3_8b_*.safetensors	none	flux2-vae.safetensors

.gguf diffusion models and text encoders show up in the same dropdowns when ComfyUI-GGUF is installed.

🔗 Bawk LoRA Stack

Six slots per node, each with LoRA, strength and trigger words. Trigger words are added to the prompt only while that LoRA is active (strength ≠ 0), so switching sets also switches the words. Chain stacks through lora_stack to combine sets; enabled turns a whole set off. Outputs lora_triggers and lora_info strings too.

🔀 Bawk LoRA Switch

Up to four sets in (set_a … set_d), one active dropdown. Only the chosen set is loaded. Feed the result into the encoder's lora_stack input.

🎲 Bawk Wildcard Encoder

• Applies the incoming lora_stack (from Bawk LoRA Stack / Switch) to MODEL and CLIP.
• trigger_words: prepend / append / off. Words already present in the prompt are not repeated.
• {red|green|blue} wildcards, seeded by wildcard_seed (nesting allowed).
• Outputs the resolved prompt and a LORA_INFO summary (name@strength, …) for the saver.

🐓 Bawk Sampler

Latent, sampling and VAE decode in one node. Detects FLUX.1 vs FLUX.2 from the model and picks the right latent layout and schedule.

Input	Notes
resolution	Presets, Match input / reference image, or Custom. A linked WIDTHxHEIGHT string also works.
steps / guidance	FLUX.1 dev: 20 / 3.5 · Kontext: 20 / 2.5 · FLUX.2 dev: 20 / 4.0 · klein distilled: 4 (guidance ignored)
scheduler	auto = simple for FLUX.1, official FLUX.2 schedule for FLUX.2
input_image + denoise	img2img
reference_image / reference_image_2	Kontext and FLUX.2 / klein image editing. Images are scaled to the model's preferred size and attached as reference latents.
cfg	Leave at 1.0 for all distilled models. Klein base models use real CFG (~4-5) against an empty negative.
max_shift / base_shift	FLUX.1 only

💾 Bawk Image Saver

Saves to output/[model]-DD-MM-YYYY/. PNG files embed the workflow (drag them back into ComfyUI); jpg / webp get a .json sidecar. Optional prompt .txt (with the LoRA list) and Discord webhook upload.

📸 Bawk Image Loader · 📁 Bawk Batch Processor

Load Image with resize / pad-to-square and size outputs. Batch Processor reads one row of a CSV / JSON file (prompt, seed, steps, guidance, resolution) so batch_index on increment walks through a list of prompts.

Installation

ComfyUI Manager: search for "Bawk Nodes" and install.

Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/juddisjudd/ComfyUI-BawkNodes.git

Requires a ComfyUI build with FLUX.2 support (late 2025 or newer). GGUF loading needs ComfyUI-GGUF.

Example workflows for every supported model family are in workflows/: FLUX.1 dev text-to-image, LoRA sets with trigger words + switch (flux1_dev_lora_sets.json), img2img, Kontext edit, FLUX.2 dev, klein 4B text-to-image and klein 9B two-reference edit.

Changes in 3.1.0

• LoRA trigger words per slot, Bawk LoRA Switch (A/B/C/D), LORA_INFO output and saver input, 4:5 / 5:4 presets.

Changes in 3.0.0

• FLUX.2 dev and klein support (latent layout, schedule, single text encoder), GGUF loading, reference images for Kontext / klein edit.
• Fixed: guidance was never applied (wrong conditioning key), PNG metadata was not ComfyUI-loadable, image loader validation was disabled.
• LoRA sets: new Bawk LoRA Stack (chainable, toggleable) replaces the broken dynamic stack + apply pair.
• Removed Z-Image loader (use Bawk Model Loader), the placeholder ControlNet node, the unused AI prompt enhancement code and ~1,200 lines of dead utilities.
• Widget layouts changed, so workflows saved with 2.x need their Bawk nodes re-added.

License

GPL-3.0 - see LICENSE.