Pfannkuchensacks Workflow Package

A collection of composable workflows for InvokeAI. Installing this node pack imports all of them into your workflow library automatically.

Install

Node Manager → Install Node Pack → Git Repository URL → paste this repo's URL.

All 25 workflows land in your library, tagged node-pack:Pfannkuchensack-Workflow-Package:

The pack ships one trivial node (Pfannis Dummy Node, a string passthrough). It exists because InvokeAI only accepts a node pack that contains an importable __init__.py — the workflows are the actual content.

One-time setup after installing

Open each MAIN - … workflow once and pick the sub workflow on every Call Saved Workflow node (Pass 1, Pass 2, Detail, …), then save.

Picking the workflow in the Workflow Id field makes that sub workflow's model fields appear on the call node. Set all of them, not just Transformer: VAE and the text encoder are marked optional, but leaving them empty falls back to whatever the selected checkpoint happens to carry — and a checkpoint that ships neither will fail at run time rather than at connect time.

What you never have to touch is the wiring. The connections into the call node — prompt, width, height, seed, steps, CFG — ship inside the MAIN file and survive both the import and a later change of sub workflow. Note how Pass 2 carries two inputs Pass 1 does not: Image and Denoising Start.

Importing always assigns fresh workflow IDs — InvokeAI drops the id from the file and generates a new UUID per workflow — so the sub workflow a call node points at cannot be shipped in the file. Until you pick one, the call node has no workflow selected.

What it does not cost you is the wiring: the field templates for the exposed inputs ship inside the MAIN files, so all connections into the call nodes survive the import and stay put when you pick the sub workflow. You only choose the workflow; you never rewire.

What you get

Four top-level workflows you run directly:

Workflow	What it does
MAIN - 2-Pass Upscale	Generate at a base resolution, then upscale + refine. Each pass picks its own model type.
MAIN - Upscale an existing Image	Feed in an image, upscale + refine it. Target size is derived from the image.
MAIN - Detail a part of an Image	Name a body part, re-render just that region at higher effective resolution.
MAIN - Detail two parts of an Image	Same, twice in sequence — e.g. face, then hands.

…and 21 callable sub workflows, three per model base, used by the MAINs through the Call Saved Workflow node:

	SDXL	FLUX	Z-Image	Krea-2	Anima	Klein 9B	Klein 4B
TXT2IMG - X	✓	✓	✓	✓	✓	✓	✓
UPSCALE REFINE - X	✓	✓	✓	✓	✓	✓	✓
DETAIL - X	✓	✓	✓	✓	✓	✓	✓

Each sub workflow returns image, latents and metadata.

The MAINs stay small on purpose: they hold the exposed settings and two call nodes, while the model-specific work lives in the sub workflows.

The point of the split

Every sub workflow of the same role exposes the same inputs under the same node ids. Switching the workflow on a Call Saved Workflow node therefore keeps the existing wiring — you swap SDXL for Krea-2 with one dropdown, without rebuilding the graph.

That only holds within a role. Putting a TXT2IMG workflow into a slot that expects an UPSCALE REFINE drops the image and denoising_start connections, because those fields do not exist there — and switching back does not restore them. Keep Pass 1 on TXT2IMG - … and Pass 2 on UPSCALE REFINE - ….

Settings that matter

Steps and CFG are model-specific, which is why the MAINs expose them per pass:

Model	Steps	CFG / Guidance
SDXL	~30	5–8
Krea-2-Turbo	8	1.0 (negative prompt is ignored)
Krea-2-Raw	~28	~4.5
FLUX	~25	guidance ~3.5
Z-Image-Turbo	~8	1.0

Keep resolutions a multiple of 16. Several denoise nodes declare multiple_of=16, and the 2×2 patch packing used by the Qwen-Image VAE (Krea-2, Anima) requires it.

Notes

• The model is selected inside each sub workflow and appears on the call node once a workflow is picked.
• DETAIL - … uses Grounding DINO + Segment Anything. "Detect" is a free text prompt, so anything the detector can name works — not just body parts.
• Mixing model types across passes costs one model load per switch. Staying on one base for both passes is considerably faster.
• MAIN - … workflows have no workflow_return node, so they are correctly reported as not callable from another workflow. They are the top of the chain.