🚀 Smart Ultimate SD Upscale (Custom Sample)

This is the sibling of ArchAi3D_Smart_Ultimate_SD_Upscale - same tiled upscaler, same per-tile CONDITIONING_LIST trick, same USDU grid math and seam-fix modes. The difference is in the name: it opens the hood on the sampling step. Where the plain node only lets you pick from the built-in sampler_name / scheduler dropdowns, this one accepts optional SAMPLER and SIGMAS inputs.

When you actually need it

The honest answer: most people don't. If you're running a normal KSampler with euler/karras, grab the base node. This variant exists for the workflows where the sampler itself is the experiment. A few real cases:

• You're using a custom sampler from a sampler pack (some don't register in the standard KSampler.SAMPLERS list).
• You have a pre-computed sigma schedule you want to force - one-step or few-step models, DMD/consistency-style checkpoints, or a schedule you built with a sigmas node.
• You're A/B testing sampling strategies on the same tiled upscale and want to swap samplers without rebuilding the graph.

Under the hood it's literally the base class with upscale_model moved to optional (you can skip the pixel upscaler entirely and go straight from the source image) plus the two extra optional inputs. Everything else - denoise, upscale_by, tile_width, tile_height, tile_padding, the four seam-fix modes, force_uniform_tiles, tiled_decode - behaves identically.

Inputs worth knowing

• custom_sampler - a SAMPLER object. Leave unconnected and it falls back to sampler_name.
• custom_sigmas - a SIGMAS tensor for a hand-rolled schedule. If you feed sigmas, most sampler implementations take over the step count, so set steps accordingly or it just gets ignored.
• upscale_model - now optional, so a pure latent-upscale pass is possible without an ESRGAN model.
• conditionings - still the CONDITIONING_LIST; tile order is still row-major to match the pack's Smart Tile prompters.

Output is the stitched IMAGE, same as the base node.

Gotchas

Because this is a port of the A1111 Ultimate SD Upscale script, it inherits that script's assumptions: custom_sigmas is passed into the stable-diffusion processing object the way the original script threaded custom_sampler/custom_sigmas through. If your custom sampler expects different latent handling (some of the fancier ones do), the tiles may look subtly different from a normal KSampler run of the same settings - worth a test tile before committing to a long render.

One other thing to remember: this is a "more detail" upscaler, not a restorer. Feed it a genuinely damaged source and it'll faithfully amplify the damage. For that job you want a restoration model first, this after.

Install

Same pack as everything else in the ArchAi3D family:

cd ComfyUI/custom_nodes
git clone https://github.com/amir84ferdos/ComfyUI-ArchAi3d-Qwen.git
cd ComfyUI-ArchAi3d-Qwen && pip install -r requirements.txt

Or ComfyUI Manager → search "ArchAi3d Qwen" → Install → restart. The node lives under ArchAi3d/Upscaling/USDU in the node menu. And yes, the pack's dual license still applies - free personal, paid commercial.