Mastering ComfyUI — Part 6: Upscaling & Post-Processing

NOTE

Goal for this part: get higher-resolution, higher-detail final images out of ComfyUI without just cranking up your base generation resolution (which usually makes composition worse, not better).

Previous: Part 5 — Advanced Conditioning

Why you don't just generate at high resolution directly

Every checkpoint has a native resolution range it was trained on — SDXL-family models around 1024×1024 (or equivalent total pixel counts at other aspect ratios), Flux similarly. Push Empty Latent Image well past that range and you don't get a sharper version of the same image — you tend to get duplicated subjects, warped anatomy, or repeated background elements, because the model is being asked to compose at a scale it never learned.

The standard solution is to generate at a resolution the model actually understands, then upscale afterward as a separate step. This also happens to be more efficient: a full generation at 2x resolution costs far more compute than generating at native resolution and upscaling the result.

Two different kinds of upscaling

Image upscaling (post-decode) — happens after VAE Decode, on the actual pixel image. Uses a dedicated upscale model (ESRGAN-family models are the most common) loaded via Load Upscale Model, then applied with an Upscale Image (using model) node. This is a pure image-processing step — it sharpens and enlarges based on what the upscale model learned about turning low-res detail into high-res detail, but it can't fix compositional problems, because the model has already been decoded into final pixels.

Latent upscaling (pre-decode, "hires fix") — happens before VAE Decode, directly on the sampler's latent output. You resize the latent, then run it through a second KSampler pass at a low denoise value (typically 0.3–0.5). Because this second pass is still a real (partial) diffusion pass, it can add and refine detail rather than just interpolating pixels — sharper texture, cleaner fine detail, better-defined edges — while the low denoise keeps the overall composition intact from the first pass.

TIP

Which one should you actually use? Latent upscaling ("hires fix") gives noticeably better detail quality because it's genuinely regenerating detail, not just enlarging pixels. Image upscaling is simpler, faster, and safer when you just want a bigger file without touching content — e.g., preparing a final image for print or a video pipeline that wants a specific resolution. Many workflows use both: latent upscale first for quality, then a light image upscale afterward if you need a specific final pixel size.

Building the hires-fix pattern

This inserts between your existing KSampler and VAE Decode:

Your first KSampler runs as normal (this is your existing generation, e.g. 1024×1024) and outputs LATENT.
Search and place "Upscale Latent" (sometimes labeled LatentUpscale or LatentUpscaleBy). Wire the first KSampler's LATENT output into it, and set your target scale — 1.5x–2x is the usual range. Going much higher in one step tends to reintroduce the same composition problems you're trying to avoid; if you want a bigger jump, chain two smaller upscale passes instead of one large one.
Place a second KSampler. Wire the upscaled latent into its latent_image input, and reconnect the same model, positive, and negative inputs from your first pass.
Set this second KSampler's denoise low — 0.3 to 0.5. This is the setting that makes the whole technique work: high enough to actually add detail, low enough that it doesn't repaint the image and drift away from your first pass's composition. A seed matching or differing from your first pass both work here — it doesn't need to match, since you're refining, not reproducing.
Wire the second KSampler's LATENT output into VAE Decode, same as any single-pass workflow.

WARN

Denoise too high on the hires pass undoes the point. If you push the second pass's denoise toward 0.7+, you're not refining detail anymore — you're substantially regenerating the image at the new resolution, which reopens the same composition risks (duplicated subjects, warped anatomy) that hires-fix exists to avoid. If you see a second pass changing things you didn't want changed, drop denoise before touching anything else.

Adding image-model upscaling

If you want pixel-level upscaling too — either instead of, or after, a hires-fix pass:

Download an upscale model (.pth files — ESRGAN variants like 4x-UltraSharp are widely used for photorealistic content; illustration-oriented content often pairs better with an upscale model trained on that kind of material) into models/upscale_models/.
Place Load Upscale Model and select it from the dropdown.
Place Upscale Image (using model), wire Load Upscale Model's UPSCALE_MODEL output into it, and your decoded IMAGE into its image input.
Its output feeds Save Image as usual.

This node does a fixed multiplier based on what the model was trained for (commonly 4x) — if you need a specific final resolution rather than whatever multiple that produces, follow it with a plain "Image Scale" (or "Resize Image") node set to your exact target dimensions.

Troubleshooting

BUG

Hires-fix pass makes the image worse, not better. Almost always denoise set too high on the second KSampler. Drop it to 0.3–0.4 and re-test before changing anything else.

BUG

Upscale Image (using model) node errors about mismatched channels or sizes. Check that the image feeding it actually came through VAE Decode first — this node expects a standard decoded IMAGE, not a raw latent. A LATENT-to-IMAGE mismatch is the most common cause of this specific error.

BUG

Final output looks oversharpened or has odd texture artifacts. Some upscale models oversharpen fine texture at full strength. If your upscale node has a strength/blend parameter, back it off; otherwise, follow it with a very light blur or a second, gentler upscale model — or reduce reliance on image upscaling and lean more on the latent hires-fix pass instead, since it produces more natural detail.

Next: Part 7 — Custom Nodes & Extensions.

Previous: Part 5 — Advanced Conditioning