Donut Tiled Upscale

The wall every ComfyUI user hits eventually: your image looks great at 1024×1024, but at 4K the GPU chokes, and the moment you try to denoise a big latent you get CUDA out of memory. The standard escape is tiled upscaling - process the image in overlapping tiles, blend the seams, and never hold the whole thing in VRAM at once. DonutTiledUpscale is that idea, but deliberately model-agnostic: it samples with the same approach as the core KSampler, so it works with SDXL, Z-Image/Lumina2, and the rest of the modern architectures rather than being hard-wired to one UNet layout.

How it works

Two passes, the standard shape for this kind of node:

1. Pixel upscale - your upscale_model (an ESRGAN-style model like 4x-UltraSharp, or whatever you load into an Upscale Model loader) resizes the image, tiled internally with overlap so the upscaler itself doesn't blow memory.
2. Diffusion refinement - the upscaled canvas is sliced into roughly 1-megapixel tiles, each run through an img2img denoise with your prompt, then stitched back with feathered overlap so seams don't show. feather (default 15, as a % of tile size) is your seam-blending knob.

That two-stage flow matters. The pixel pass adds the pixels; the diffusion pass adds the detail - this is "more detail" territory, not just "more pixels," so expect the model to gently re-render things. Keep your denoise low (default 0.35) unless you want a rewrite.

Inputs that actually matter

The plumbing is standard: image, upscale_model, model, positive/negative, vae - then sampler settings (seed, steps, cfg, sampler_name, scheduler, denoise) exactly like a KSampler. The Donut-specific ones:

• rescale_factor - the upscale factor (1–8, default 2). The tooltip explains the real behavior: regular mode picks ~1MP tiles; edit mode snaps the target to a 32px grid.
• resampling_method - lanczos (default), bicubic, bilinear, or nearest. Lanczos is the right default.
• feather - seam blend width, regular mode only.
• tiled_vae - use a tiled VAE decode to save even more VRAM (regular mode only).
• tiled_diffusion (optional, default on) - process in overlapping diffusion tiles. Turn it off for one full-resolution diffusion pass; the tooltip is not kidding about needing "substantially more VRAM."
• turbo_mode (optional) - for distilled models (Z-Image Turbo, Lightning): treats your steps as the model's supported Turbo steps and snaps denoise to the nearest valid scheduler point. Use this if your images come out over-sampled on a Turbo model.

Outputs are image and debug_image - the second shows the tile layout, which is genuinely useful the first time seams appear.

The edit mode

There's an optional edit_mode that's a completely different animal: one full-frame Krea2 img2img upscale against a full reference, no diffusion or VAE tiling. It needs clip (required in this mode), an edit_source_image identity reference, edit_prompt/edit_negative_prompt, and optionally edit_model (a Krea2 model with the Identity Edit LoRA pre-applied). If that sentence means nothing to you, ignore the toggle - it's for identity-preserving edits on Krea2 models, and the default off is the sensible place for it.

Install

Ships in ComfyUI-DonutNodes: ComfyUI Manager → search "DonutNodes" → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/DonutsDelivery/ComfyUI-DonutNodes.git donutnodes
cd donutnodes
python -m pip install -r requirements.txt

No model downloads in the pack itself - but the node needs an upscale_model you supply (any ESRGAN-style .safetensors in your upscale_models folder). Dependencies are opencv-python-headless, scipy, matplotlib, psutil, tqdm, requests.

Where people get burned

• Seams or tile-boundary artifacts - raise feather, and glance at debug_image to see your actual grid.
• Content drift at high denoise - this is a detail pass, not a free lunch. Denoise above ~0.5 and the model starts re-imagining; below ~0.2 and it does almost nothing. 0.3–0.4 is the honest sweet spot for most models.
• Edit mode with no CLIP - it hard-requires clip and an edit_source_image; wire them or the mode is a footgun.
• VRAM still OOMs - that's tiled_diffusion off or tiled_vae off. Turn both on for the memory-friendly path.

Compared to the established giants (Ultimate SD Upscale from Impact Pack, Tiled Diffusion), this doesn't reinvent the wheel - its selling point is that it samples like core ComfyUI, so it keeps working as new architectures (Z-Image and friends) arrive. If you're on SDXL and already happy with Ultimate SD Upscale, there's no urgent reason to switch. If you're on a newer model where the old tiling nodes misbehave, this is the one to reach for.