HSWQ Ultimate SD Upscale

Ultimate SD Upscale is the tiled-upscaling workhorse that got big in A1111 and carried over to ComfyUI: upscale the image with a pixel upscaler, then run a diffusion img2img pass on overlapping tiles so the model adds real detail without ever seeing the whole huge image at once - which is how you get 4K output on a 12 GB card. This node is that, shipped inside the HSWQ pack as a self-contained copy, with the original ComfyUI_UltimateSDUpscale package deliberately not required at runtime. It's based on ssitu's GPL-3.0 project, and the README says so openly.

What's actually different here

• Color normalization. Nunchaku SDXL VAE output comes back in a compressed range (roughly 0.15–0.85), which reads as pale, washed-out color. This node normalizes VAE output to full 0.0–1.0 range before upscaling, fixing that specific failure out of the box.
• FP8/torch.compile compat patches. On load it applies patches for copy shape mismatches, FP8 linear/addmm bias–out_features mismatches, and control embedder weight layout issues, so the node actually works with HSWQ FP8 models and torch.compile.
• Standalone. No separate install of ssitu's pack, no module-name collisions with other custom nodes.

How the upscaling is sized

• upscale_by - the big one. Auto reads the input image height and sets scale = target_height / input_height, clamped to 0.05–4.0. Or pick a fixed magnification from the dropdown (0.05 to 4.00 in 0.05 steps) and target_height is ignored.
• Example straight from the README: input height 1080, upscale_by = Auto, target_height = 4320 → scale 4.0 → output 4320px tall.

The rest is the standard Ultimate SD Upscale kit: upscale_model (an ESRGAN-style pixel upscaler), tile_width / tile_height (512 default), mask_blur, tile_padding, mode_type (Linear / Chess / None - tile order), the seam-fix family (None / Band Pass / Half Tile / Half Tile + Intersections with seam_fix_denoise, seam_fix_width, seam_fix_mask_blur, seam_fix_padding), plus the sampling set - seed, steps, cfg, sampler_name, scheduler, denoise (0.2 default, the classic light-touch tile value), positive, negative, vae, model, image. Two settings deserve special attention:

• batch_size - defaults to 1. The tooltip is a trap warning: higher values are faster but eat VRAM, and if you get tensor-size mismatches with FP8/FP4 quantized models, set this to 1. It's the first thing to check when a tiled run starts erroring on quantized weights.
• tensor_boost - off by default, and the README's guidance is to keep it off for tiled upscaling. Turning it on enables Per-Weight CUDA Graph acceleration for SDXL ConvRot NVFP4 on Blackwell, and it costs several GB of VRAM in CUDA Graph arenas. The intended pairing is: Tensor Boost on for the fixed-resolution base pass in the HSWQ Sampler, off here for the tiles, because per-tile shape changes stack graph arenas and can spill to system RAM. Recommended hardware for that path: RTX 5090 with 32 GB+.

Gotchas

Beyond the batch_size and tensor_boost notes, remember the pack's standing rule: if you're upscaling HSWQ ConvRot INT8/NVFP4 models, put General Purge VRAM V2 from ComfyUI-DistorchMemoryManager at the end of the workflow with the HSWQ toggle on, or the second run can fail on residual NVFP4 pools. And if you're not running the HSWQ stack at all, ssitu's standalone Ultimate SD Upscale is the more widely documented choice - this node's edge cases (color normalization, FP8 patches) only matter to this ecosystem.

Install: ComfyUI Manager or git clone https://github.com/ussoewwin/ComfyUI-HSWQ-Loader-and-Tools into custom_nodes, then restart.