DeepSeek-V3 Demonstrates Superior Handling of Complex ComfyUI JSON Logic for High-Resolution Upscaling

The process of developing a high-resolution pipeline started with DeepSeek-V3, which shows clear advantages over GPT-4o when dealing with intricate ComfyUI JSON structures. Pushing pixel density beyond 4K using the Ultimate SD Upscale node reveals that the bottleneck is not solely VRAM-related but also heavily influenced by denoising thresholds and tile seam management.

During a recent evaluation period, side-by-side benchmarks between Claude 3.5 Sonnet and DeepSeek-V3 were conducted to determine the optimal prompts and node configurations for eliminating the unwanted "grid lines" in tiled upscaling results. Claude 3.5 Sonnet excels in creating descriptive aesthetic prompts, but DeepSeek-V3 proves more effective in managing the technical intricacies of ComfyUI's latent space logic.

Methods to Remove Tile Seams in Upscaling

The solution to eliminating seams involves carefully adjusting the upscale_by ratio in relation to the denoise value. Setting denoise above 0.35 introduces ghosting effects, whereas a value below 0.2 fails to enhance detail. The best balance found was 0.3 for a 2x upscale, though a critical step is incorporating a Model Sampling Discrete node to stabilize noise.

Analysis of Model-Driven Prompting for Upscaling

Performance Analysis: Model-Driven Prompting for High-Resolution Upscaling

Claude 3.5 Sonnet
Advantages: Highly effective at generating precise "detail injection" prompts. If the goal is to add realistic skin textures or fabric patterns without distorting the original composition, Claude's prompts are highly targeted.
Limitations: Tends to produce overly verbose outputs, which can clutter the prompt field and diminish the impact of LoRAs.

DeepSeek-V3
Advantages: Exceptional at calculating optimal tile dimensions and VRAM usage. For a system configured with a 24GB 3090, DeepSeek-V3 correctly recommended a 512x512 tile size with a 64px overlap to reduce seams while maintaining high throughput.
Limitations: Sometimes neglects the overall artistic tone of the image, prioritizing technical precision over creative expression.

GPT-4o
Advantages: Reliable for basic troubleshooting and explaining node functions.
Limitations: Often recommends outdated nodes or generic settings that do not account for recent ComfyUI updates.

When configuring the Ultimate SD Upscale node, avoid the default "Chess" seam blending method. Instead, switch to "Random" or "Spiral" and apply the following prompt structure to ensure consistency across tiles:

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

Optimizing Node Settings and Workflow Efficiency

To improve speed, a "latent upscale" is now performed first, followed by the Ultimate SD Upscale node for the final adjustment. This approach reduces tiling time by approximately 30%. On the tested setup, a standard tiled 4K upscale took 4 minutes, but pre-scaling the latent space and applying a lower denoise in the final pass cut the time to 2 minutes and 15 seconds with minimal quality degradation.

The most crucial setting is the upscale_by multiplier. Incremental scaling is preferred over direct jumps to 4x. Instead, two passes of 2x are recommended. The first pass handles the initial structural expansion, while the second pass—using a specialized, detail-oriented model—addresses artifacts. This staged upscaling method is the only way to avoid the "plastic" appearance that results from excessive single-pass upscaling.