Optimizing ComfyUI Workflow for High-Resolution Upscaling Using Ultimate SD Upscale

DeepSeek-V3 handles the logic for complex ComfyUI JSON structures far better than GPT-4o, which is where I started my testing for this high-res pipeline. When you're pushing pixels past 4K using the Ultimate SD Upscale node, the bottleneck isn't just VRAM—it's the denoising threshold and the tile seam management.

I spent the weekend running head-to-head benchmarks between Claude 3.5 Sonnet and DeepSeek-V3 to generate the specific prompts and node configurations needed to eliminate those dreaded "grid lines" in tiled upscales. Claude is great for descriptive aesthetics, but DeepSeek is a beast at the technical side of ComfyUI's latent space logic.

For those struggling with seams, the secret is in the upscale_by ratio versus the denoise value. If you push denoise above 0.35, you get ghosting; below 0.2, you get no detail enhancement. I found the sweet spot at 0.3 with a 2x upscale, but the real trick is using a Model Sampling Discrete node to stabilize the noise.

Performance Breakdown: Model-Driven Prompting for Upscaling

Claude 3.5 Sonnet
Pros: Incredible at "detail injection" prompts. If you need the upscale to add realistic skin pores or fabric weave without changing the composition, Claude's prompts are surgical.
Cons: Tends to be overly verbose, which can sometimes clutter the prompt field and dilute the impact of the LoRAs.

DeepSeek-V3
Pros: Superior at calculating tile sizes and VRAM overhead. I asked it to optimize my workflow for a 24GB 3090, and it correctly suggested a tile size of 512x512 with a 64px overlap to minimize seams while maximizing throughput.
Cons: Occasionally misses the "vibe" of the image, focusing too much on technical accuracy over artistic flair.

GPT-4o
Pros: Reliable for basic troubleshooting and explaining what a specific node does.
Cons: Often suggests outdated nodes or generic settings that don't account for the latest ComfyUI updates.

If you're setting up your Ultimate SD Upscale node, stop using the default "Chess" seam blending. Switch to "Random" or "Spiral" and use this specific prompt structure to maintain consistency across tiles:

(high resolution:1.2), (extremely detailed:1.2), 8k uhd, highly detailed textures, sharp focus, [insert specific subject], cinematic lighting, intricate details

To truly optimize for speed, I've been using a "latent upscale" first, followed by the Ultimate SD Upscale node for the final pass. This reduces the time spent in the tiling phase by about 30%. On my rig, a 4K upscale using a standard tiled approach took 4 minutes; by pre-scaling the latent and using a lower denoise on the final tile pass, I got it down to 2 minutes and 15 seconds with virtually no loss in quality.

The most critical setting is the upscale_by multiplier. Don't jump straight to 4x. Do two passes of 2x. The first pass handles the structural enlargement, and the second pass—using a different, more detail-oriented model—cleans up the artifacts. This "staged upscaling" is the only way to avoid the "plastic" look that happens when you push a single-pass upscale too far.