Learn › ComfyUI Advanced: Training, Kontext, and Automation › Ultimate SD Upscale: Tiled High-Resolution Output

Ultimate SD Upscale: Tiled High-Resolution Output

Advanced 🕐 11 min Lesson 4 of 12
What you'll learn
  • Install the Ultimate SD Upscale custom node via ComfyUI-Manager
  • Connect the UltimateSDUpscale node correctly within a post-ESRGAN upscaling workflow
  • Set tile_width, tile_height, denoise, and seam_fix_mode appropriately for FLUX and SDXL outputs
  • Explain when to use Ultimate SD Upscale versus the simpler detail-pass approach from ComfyUI Image Generation

How the Detail Pass Falls Short at Very High Resolution

The three-stage upscaling pipeline from ComfyUI Image Generation — generate at base resolution, ESRGAN 4×, then a low-denoise KSampler pass on the whole image — produces good results at 2–3K output sizes. The detail pass re-runs the diffusion model on the already-upscaled image, adding fine texture and sharpness that the upscale network alone cannot generate.

At 4K and above, the detail pass starts to break down. The problem is VRAM: running the diffusion model on a 4096×4096 image in a single pass requires more VRAM than most consumer GPUs can hold. Even with tiled VAE enabled, the KSampler itself processes the full latent at once, which at 4K is roughly 512×512 in latent space. Many cards simply run out of memory before the pass completes.

The second problem is consistency. A single-pass detail run at high resolution tends to sharpen the entire image uniformly, which can look over-processed — areas of intentional softness get sharpened, gradients gain unwanted grain. What you actually want is targeted, locally-appropriate detail added separately to each part of the image.

Ultimate SD Upscale solves both problems by tiling. It splits the high-resolution image into tiles the size of your model's native resolution, runs the full diffusion pipeline on each tile independently, and stitches them back together with a seam-fix pass to eliminate grid artifacts. Each tile gets the detail treatment it needs, the VRAM cost stays constant regardless of output size, and the final image has genuine AI-generated detail everywhere — not just where a single large pass managed to reach.

Installing Ultimate SD Upscale

Ultimate SD Upscale is a custom node. It is not included with ComfyUI by default and must be installed via ComfyUI-Manager.

Open ComfyUI-Manager (the Manager button in the top right of the interface), go to Install Custom Nodes, and search for UltimateSDUpscale. Install the result by ssitu. After installation, restart ComfyUI. The node appears in the node menu under image/upscaling as Ultimate SD Upscale.

No additional model files are required. The node uses your existing diffusion model, text encoders, VAE, and upscale model — whatever is already in your workflow.

Connecting the Node

The Ultimate SD Upscale node is a drop-in replacement for the KSampler + VAE Decode stage at the end of a standard upscaling workflow. Here is how it fits into the pipeline:

  1. Run your generation as normal. At the end, VAE Decode and Save Image give you the base result.
  2. Add a Load Upscale Model node and an Upscale Image With Model node (covered in ComfyUI Image Generation). Connect your base image through the ESRGAN upscale to get the 4× enlarged version.
  3. Connect the ESRGAN output to the Ultimate SD Upscale node's image input. Also connect your model, positive conditioning, negative conditioning (or empty conditioning for FLUX), VAE, and upscale model.
  4. Connect the Ultimate SD Upscale output directly to a Save Image node.

The node handles VAE encoding and decoding internally — you do not need separate Encode and Decode nodes around it. The output is already decoded pixels, ready to save.

The key inputs that need connecting:

  • image: the ESRGAN-upscaled image (not the original base-resolution image)
  • model: your loaded checkpoint or diffusion model
  • positive / negative: your CLIP Text Encode outputs (use the same prompt as the original generation)
  • vae: your loaded VAE
  • upscale_model: your ESRGAN model (for a second-stage upscale within the tile pass — often set to None if you have already upscaled to your target resolution)
  • seed, steps, cfg, sampler_name, scheduler: same values as your generation workflow

The Key Parameters

Ultimate SD Upscale has more settings than a standard KSampler. These are the ones that matter most:

tile_width and tile_height: The dimensions of each tile the image is split into. This should match your model's native resolution — 1024×1024 for SDXL and FLUX, 512×512 for SD 1.5. Using the wrong tile size produces tiles that look mismatched from each other, since the model was trained at a specific resolution. Default is 512; change it to 1024 for SDXL and FLUX.

denoise: How much the diffusion process changes each tile. Range 0.0 to 1.0. The practical range is 0.2–0.5:

  • 0.2–0.3: subtle sharpening and micro-detail, minimum change to composition
  • 0.3–0.4: visible detail enhancement, slight texture variation — the most common starting point
  • 0.4–0.5: noticeable re-texturing, may change specific areas more than expected
  • Above 0.5: significant changes to the image content, not just detail — rarely appropriate for upscaling

Start at 0.3 and adjust based on your results. For images where you want to add significant new detail (textures, fabric, foliage), move toward 0.4–0.45. For clean renders or portraits where you want precision without alteration, stay at 0.2–0.3.

seam_fix_mode: Controls the stitching pass that eliminates visible tile boundaries. Options are None, Band Pass, Half Tile, and Half Tile + Intersections. Half Tile is the recommended default — it runs a second pass over the border area of each tile pair, blending the seam. For large tiles at high denoise, switch to Half Tile + Intersections which also processes the corners where four tiles meet. Use None only for testing or when render time is a priority.

mode (redraw_mode): Linear processes tiles left-to-right, top-to-bottom. Chess processes tiles in a checkerboard order, which can produce slightly more consistent results across the image because adjacent tiles are processed alternately rather than sequentially. Chess is generally preferred for high-quality output; Linear is faster.

tile_padding: The number of pixels of overlap between tiles. Default 32. Increasing to 64 can help reduce seam visibility at higher denoise values, at the cost of longer processing time.

The Complete High-Resolution Pipeline

Putting it all together, the recommended workflow for 4K output from a 1024px FLUX generation:

  1. Generate at 1024×1024 (or 1024×768). Get the composition right here. Do not proceed to upscaling until the base result is what you want.
  2. ESRGAN 4× using Load Upscale Model + Upscale Image With Model with 4x-UltraSharp. Your 1024px image is now 4096px.
  3. Ultimate SD Upscale at denoise 0.3, tile_width/height 1024, seam_fix_mode Half Tile, Chess mode. This adds diffusion-generated detail to every tile of the 4K image.
  4. Save the final output. The result has genuine model-generated detail across the full 4K canvas — hair strands, fabric weave, surface texture — that neither ESRGAN alone nor a single-pass detail run can match.

Approximate render times on an RTX 4090 at 4K with 20 steps and Chess mode: 8–12 minutes. On an RTX 3080 (10GB): 20–30 minutes. If render time is a constraint, use Linear mode and reduce steps to 15 — the quality difference is minor and the time saving is significant.

When to Use Ultimate SD Upscale vs the Simple Detail Pass

Both approaches are valid. The simple detail pass (from ComfyUI Image Generation) is faster and requires no extra nodes. Ultimate SD Upscale is slower but produces higher quality at large output sizes. Here is when to use each:

  • Use the simple detail pass for output sizes up to roughly 2K (2048×2048), when render time matters, for quick iteration, or when the ESRGAN output already looks good and you just need a final sharpening pass.
  • Use Ultimate SD Upscale for output at 3K and above, for print-quality files where fine texture everywhere matters, when the simple detail pass runs out of VRAM, or when you need consistent detail quality across the whole canvas rather than a uniform sharpening.

The two approaches can also be combined: run ESRGAN → simple detail pass to get a clean 2K intermediate, then Ultimate SD Upscale on the 2K result to reach 4K or 8K. This two-pass approach is slower but can produce particularly detailed results for large prints.

Key takeaways
  • Ultimate SD Upscale splits the image into model-native-sized tiles and runs the full diffusion pipeline on each one — this keeps VRAM usage constant regardless of output size and adds genuine AI detail everywhere
  • tile_width and tile_height should match the model's native resolution: 1024 for SDXL and FLUX, 512 for SD 1.5 — using the wrong size produces visually mismatched tiles
  • Start denoise at 0.3 and adjust: 0.2-0.3 for subtle sharpening, 0.3-0.4 for visible texture, above 0.5 starts changing image content rather than just adding detail
  • Half Tile seam_fix_mode blends the borders between tiles and should be the default; switch to Half Tile + Intersections at high denoise values or large tile sizes
  • The simple detail pass is faster and sufficient up to 2K; Ultimate SD Upscale is the right choice for 3K+ output, print files, or when the detail pass runs out of VRAM
Learn more every week Get 5 hand-picked prompts + new AI tools every Monday. Free.