juddisjudd-ComfyUI-BawkNodes/.github/copilot-instructions.md

ComfyUI-BawkNodes

Custom node pack for ComfyUI focused only on FLUX models (FLUX.1 dev / Krea / Kontext, FLUX.2 dev / klein, GGUF builds of those). Goal: the fewest nodes needed for multi-LoRA FLUX workflows. FLUX has no negative prompt, so no node exposes one.

Layout

• __init__.py exports NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS, WEB_DIRECTORY.
• nodes/ one file per node (V1 class schema: INPUT_TYPES, RETURN_TYPES, FUNCTION, CATEGORY, DESCRIPTION).
• nodes/lora_stack.py also holds the shared LoRA helpers (normalize_stack, merge_stacks, apply_lora_stack, stack_info).
• js/bawk_nodes.js node colors and prompt textarea sizing only.
• workflows/ example workflow.

Nodes

Node id	Purpose
DiffusionModelLoader	model + VAE + text encoder(s). 2 encoders = FLUX.1 (CLIPType.FLUX), 1 encoder = FLUX.2 (CLIPType.FLUX2). .gguf files go through ComfyUI-GGUF's node classes looked up in nodes.NODE_CLASS_MAPPINGS at run time.
BawkLoraStack	6 slots (lora, strength, trigger words), chainable, enabled master switch. LORA_STACK is {"loras": [(name, strength)], "triggers": [str]}.
BawkLoraSwitch	picks one of set_a..set_d; only that set's LoRAs load.
FluxWildcardEncode	applies lora_stack, adds active trigger words (prepend/append/off), resolves `{a
BawkSampler	builds latent from model.get_model_object("latent_format") (16ch/8 for FLUX.1, 128ch/16 for FLUX.2), sets guidance in conditioning, optional reference_latents for Kontext / klein edit, cfg > 1 uses CFGGuider with zeroed negative (klein base only), FLUX.1 gets the resolution-dependent shift patch, FLUX.2 uses its own schedule when scheduler is auto.
FluxImageSaver	[model]-DD-MM-YYYY/ folders, PNG metadata written with json.dumps like core SaveImage, optional Discord webhook.
BawkImageLoader	Load Image + resize / pad + size outputs.
BawkBatchProcessor	one row of a CSV / JSON file -> prompt, seed, steps, guidance, resolution.

Conventions

• Only add inputs that matter for FLUX. Defaults follow the official Comfy templates (euler, 20 steps, guidance 3.5 / 2.5 Kontext / 4.0 FLUX.2, klein distilled 4 steps).
• Raise exceptions on failure; do not return fake fallback tensors.
• Use logging, not print. No startup banners.
• Keep comments to non-obvious constraints only.
• No eval / exec, no runtime pip install (registry rules).
• When changing widget inputs, update workflows/example_bawknodes_workflow.json widgets_values to match the new order.