Monument Two
I'm thrilled that nearly 400 of you have explored the M1 walkthrough since last week, with over 50 users downloading and testing the workflow. Your support has been incredibly encouraging.
This is, of course, merely the beginning. Today, I'm excited to introduce M2. This iteration features refined arrangements and integrates highly needed features such as ControlNets, Redux, and - what I think could be - a powerful upscaling system.
Below, I've outlined all the changes in a streamlined log format. For questions or feedback drop me a DM (linkedin, reddit) or an email at hey@zablit.ca!
PS: The JSON is available for download at the bottom of the page :)
Rearranged nodes and groups for improved clarity.
The canvas now features three distinct columns:
User controls and viewers (inputs/outputs)
Preprocessing (latent switches, PuLID, Redux, ControlNets)
Core processing (base gen, upscaling, and refinements)
While content within each group still flows left to right, the groups themselves are now organized in a strict top-to-bottom arrangement. I've eliminated shared sidewalls between groups for cleaner visual separation. (Will this carry into v3? We'll find out soon enough.)
Config group
The Config group is now (hopefully) clearer and better structured. I can at a glance figure out the sampler settings from the latent ones.
Controlnet
Important: Use only the 'Set Shakker Labs Union Controlnet Type' node. Previous implementations mixed various types, causing headaches. Avoid headaches at all cost. Comfy has enough of those already. :P
The integrated switch allows seamless toggling between different preprocessors based on your needs.
The image resize nodes automatically ensure your ControlNet dimensions match your main canvas, regardless of whether you're using img2img or txt2img. I've set the default method to 'Pad' (which expands the ControlNet canvas to fit the main dimensions), but feel free to experiment with other resize methods that might better suit your specific projects.
Redundancy switches
The 'Any Switch' node functions essentially as a smart redundancy system. It evaluates inputs sequentially—checking the first connection, and if nothing is detected, automatically advancing to the second, continuing this process until it finds an active input.
I've leveraged this functionality to create fault tolerance within the workflow. This means you can disable entire groups without disrupting the overall process. For example, if you choose not to load Redux, the switch seamlessly redirects to ControlNet and continues operation. Similarly with PuLID—disable that group, and the system defaults to using the base model instead.
A small but frankly insanely useful feature for large workflows.
Upscaler
This seriously warrants its own dedicated tutorial (coming next!), but here are the essential components:
The upscaling process is structured in 3-4 distinct stages, tailored to your requirements:
Initial Upscaler (Stage 1): Transforms base 1.5K resolution to 3K. For portraits, it reintegrates original details before facial processing.
Face Correction: Automatically detects and refines facial features through targeted inpainting at 1280×1280px resolution. (It can be safely turned off, the redundancy switch in Stage 2 will pick up on it.)
Secondary Upscaler: Elevates content to 6K while preserving critical details that might otherwise be lost during the transition.
Optional Final Stage (Stage 3): Pushes resolution to 12K (a whopping 100 megapixels).
Do note that unlike the creative upscaling found in commercial software, this pipeline is specifically designed for faithful, lossless enlargement optimized for print-quality output.
This completes our overview of M2. While there's certainly more to explore, I'll save those discussions for my social channels and upcoming practical tutorials where we'll put these features to work. I hope you find value in this workflow. I'm available if you'd like to connect or have questions about implementation.