Skip to content

Product Image HD Upscaling: Conservative Upscaling and Z-Image-Turbo 2K ​

💡 One-sentence summary: Use ComfyUI to build a product image HD upscaling pipeline — conservative upscaling for fidelity, Z-Image-Turbo 2K to restore detail, cloud nodes as a fallback.

Platform resolution requirements keep climbing, yet AI-generated product images are often only 480–720p; the official upscaling guide sums up this pipeline as "generate → repair → upscale → deliver." This lesson helps you solidify the final step of ComfyUI product image upscaling: first learn the most reliable conservative upscaling, then use the official Z-Image-Turbo 2K template to restore detail and resolution together.

What you'll be able to do after this lesson ​

  • Use two nodes (Load Upscale Model + Upscale Image (Using Model)) to conservatively upscale a product image 4x, keeping small label text undistorted
  • Load the official "Image Upscale: Z-Image-Turbo 2K" template with one click, upscaling the product image to 2K resolution while restoring detail
  • Explain what happens at each of the three denoise tiers (0.15–0.25 / 0.25–0.35 / >0.35) and which tier suits product images
  • Know the cloud alternatives: when Magnific Precise, Nano Banana Pro, and other partner nodes are worth using

Why use ComfyUI instead of online upscaling tools? ​

CapabilityOnline upscaling toolsComfyUI Local workflow
BillingCharged per image or by credits — batch new listings hurt✅ Download the model once, no per-image charges
Product image privacyImages must be uploaded to their servers✅ Everything runs on your own computer
Model choiceYou get whatever they give you✅ RealESRGAN / BSRGAN / SwinIR / diffusion upscaling, swap freely
PipelineUpload images one by one manually✅ Cutout → upscale → composite in one chain; the transparent background from the last lesson plugs right in
Getting started costJust open a webpage and go❌ You need to install ComfyUI and download models — this lesson walks you through it

Your current predicament ​

  • AI-generated product images are often only 480–720p, while the resolutions platforms and designers demand keep rising — delivering them as-is won't cut it
  • Directly enlarging the size leaves edges blurry and small text a smudged mess — the upscale accomplishes nothing
  • You've tried "one-click HD," only to have the product logo repainted and label text altered — the worst outcome for product images

When to use this technique ​

When you need to:

  • Upscale AI-generated images (commonly 480–720p) or phone photos to the resolution required by platforms or printing
  • Keep materials, label edges, and small text identical to the original after upscaling — product images must not "change appearance"
  • Upscale a batch of product images repeatedly without paying per image

And you don't want to:

  • Upload product images to an online service
  • Learn an entire diffusion model stack from scratch — conservative upscaling needs only two nodes

🎒 Preparation Before Starting ​

  • Complete the E-commerce Image Hands-On Overview and know that each lesson starts from an official template
  • ComfyUI can generate images normally and you know how to load workflows from the template library (if unfamiliar, review Template Library Tips)
  • Have a product image on hand: AI-generated or a real photo both work; the transparent-background PNG cut out in the last lesson is fine for conservative upscaling (transparency is preserved), but composite the background first before taking the 2K diffusion route
  • For conservative upscaling: reserve about 64MB of disk space (RealESRGAN_x4plus.safetensors is 63.8MB)
  • For Z-Image-Turbo 2K: ComfyUI 0.9.0 or later (the minimum version marked in the template library), plus about 19GB reserved for the trio; for VRAM, refer to the 16GB consumer-grade positioning in the Z-Image-Turbo Text-to-Image lesson
  • Your network can reach Hugging Face; if the connection is unstable, use the ModelScope domestic mirror (addresses given in steps 1 and 3)

Follow Along ​

Step 1: Download the upscaling model and put it in the upscale_models directory ​

Why: Conservative upscaling relies on a local super-resolution model file, so you must first place it where ComfyUI can find it.

The minimal setup only needs one file: RealESRGAN_x4plus.safetensors (63.8MB), downloadable from either repository below:

  • Hugging Face: Comfy-Org/Real-ESRGAN_repackaged (huggingface.co)
  • ModelScope (fast in China): Comfy-Org/Real-ESRGAN_repackaged (modelscope.cn)

Put it into ComfyUI/models/upscale_models/. If you forget the directory rules, review Where to Put Models.

Want a different model: OpenModelDB offers a large selection of super-resolution models—RealESRGAN is general-purpose, BSRGAN excels at sharp text edges, and SwinIR excels at natural textures and landscapes; downloaded .pth files also go into upscale_models.

You should see: RealESRGAN_x4plus.safetensors appears in the dropdown list of the Load Upscale Model node (after manually placing the file, you need to refresh or restart ComfyUI).

Step 2: Run conservative upscaling—two nodes upscale the product image by 4x (Method A, recommended starting point) ​

Why: The first principle of product image upscaling is "don't change the content." The official recommendation for product photography is conservative upscaling—preserving materials, label edges, and small text.

The node chain has four nodes in total:

Load Image ──→ Upscale Image (Using Model) ──→ Save Image
                    ↑
Load Upscale Model ─┘(选 RealESRGAN_x4plus.safetensors)

Create a new blank workflow, connect the four nodes as shown, load your product image via Load Image, then click Queue or press Ctrl+Enter to run.

You should see: The output image resolution is 4x the input (x4plus is a 4x model), with clear small text, clean edges, and unchanged content.

💡 This is the unique value of ComfyUI

Swap models anytime you're not satisfied: change the dropdown to BSRGAN or SwinIR and run again, comparing results directly in the same workflow—no extra cost, and your product images never leave your computer.

Step 3: Load the "Image Upscale: Z-Image-Turbo 2K" official template (Method B) ​

Why: Pure conservative upscaling offers limited improvement for poor-quality inputs. The official Z-Image-Turbo 2K template first pre-upscales with RealESRGAN, then resamples with Z-Image-Turbo at medium denoise, rebuilding detail at 2K resolution.

Open Workflow → Browse Templates, search for Image Upscale: Z-Image-Turbo 2K (shown as "Image Upscale: Z-Image-Turbo 2K" in the Chinese interface) and load it; you can also directly download the official template JSON (utility_z_image_turbo_2k_upscaler.app.json, in the Comfy-Org/workflow_templates repository on GitHub).

The template needs four model files; after downloading, put each in its place:

FileSizeDirectory to place in
z_image_turbo_bf16.safetensors11.46 GBdiffusion_models
qwen_3_4b.safetensors7.49 GBtext_encoders
ae.safetensors319.8 MBvae
RealESRGAN_x4plus.safetensors63.8 MBupscale_models

Download the first three from the Comfy-Org/z_image_turbo repository (available on both Hugging Face and ModelScope); they are the same files as in the Z-Image-Turbo Text-to-Image lesson, so reuse them if you already downloaded them.

You should see: On the canvas, Load Image → a subgraph node named "Image Upscale(Z-image-Turbo)" → Save Image, plus an ImageCompare comparison node alongside. Prompt, denoise, seed, and the four model dropdowns are all exposed directly on the subgraph node—no need to go inside the subgraph to change them.

Step 4: Upload a product image, tune parameters per the denoise table to produce a 2K image ​

Why: Denoise determines how far the model can stray from the original image—it's the lifeline parameter for product image upscaling.

Steps:

  1. Load Image Upload your product image (the official example input is a fighter jet photo fighter_jet.png, downloadable from the template's input directory)
  2. Keep the template default for the positive prompt masterpiece, 8k, or add product descriptions per the "Magic Prompt" below; leave the negative prompt empty (template default)
  3. Pick a denoise level from the table:
denoiseEffectRecommendation for product images
0.15–0.25Stays close to the original, minimal changesChoose this if you fear content drift
0.25–0.35Repaints more detailThe template default 0.33 is in this range
> 0.35Prone to artifactsDon't touch unless paired with a detailed description
  1. Run with Ctrl+Enter. Keep the remaining KSampler parameters as set by the template: 5 steps, cfg 1, dpmpp_2m_sde sampler + beta scheduler.

You should see: The template first normalizes the input to 1 megapixels (ImageScaleToTotalPixels), upscales 4x with RealESRGAN, then scales back to half with ImageScaleBy, and hands it to Z-Image-Turbo for resampling—a net effect of about 2x, outputting 2K-class resolution; the saved file carries the z-image-upscaled prefix.

Step 5: Before/after comparison acceptance check: small text, edges, materials ​

Why: Product image upscaling must be verified—"looks clearer" does not mean "unchanged."

The ImageCompare node in the template is already wired to the original and the upscaled result; drag the slider to compare. Focus on three things: label text, logo edges, and material texture.

⚠️ ImageCompare not showing?

The template notes state: Image Compare is currently only available under Nodes 2.0. Switch via Menu → Nodes 2.0 and it will work.

You should see: In the upscaled image, small text strokes are complete, label edges are sharp, and material texture is natural—only when all three match the original can the image go into your product detail page.

Method C: Tight on VRAM? Conservative upscaling in the cloud ​

When local VRAM can't run Z-Image-Turbo 2K, the official upscaling guide lists these cloud conservative options for product photography (cloud nodes are billed by Comfy account credits):

Partner NodeCharacteristics1K→4K official benchmark time
Magnific PreciseHigh precision, production-safe~40s
WaveSpeed SeedVR2High fidelity~40s
HitPawDual conservative/creative modes~60s
Nano Banana ProFast, well-suited for product photography~80s

The official upscaling guide page offers one-click entries for "Comfy Cloud trial run / download workflow" for HitPaw, Nano Banana Pro, SeedVR2, and others. Tip: for the SeedVR2 route, the official recommendation is to first use ImageScaleToTotalPixels to scale the image down to 0.35 megapixels before upscaling for better results.

📋 Magic Prompt ​

Write both of these into the positive prompt of the 2K template (the text field on the subgraph node).

📝 Short-tag version (denoise ≤ 0.33) ​

Expected result: adds clarity while staying close to the original image, suitable for most product photos.

masterpiece, 8k

This is the template default. For more specificity, append a product description afterwards, e.g. masterpiece, 8k, white ceramic coffee mug on clean background.

📋 Detailed description version (required when denoise > 0.35) ​

Expected result: keeps structure consistent at high denoise — the official template comments gave a comparison: at denoise 0.5, using masterpiece, 8k tends to fall apart, while switching to a full description is noticeably more stable.

[One full description: viewpoint + product + material color + label text and position + background lighting], masterpiece, 8k

Examples:

A front view of a white ceramic coffee mug, matte glaze, logo printed on the side, soft studio lighting, clean background, masterpiece, 8k

Z-Image-Turbo's text encoder is Qwen3 4B, which handles both Chinese and English prompts — writing the detailed description in Chinese works just as well.

Checkpoints ✅ ​

  • Load Upscale Model dropdown can select RealESRGAN_x4plus.safetensors
  • Conservative upscaling runs successfully: output resolution is 4x the input, small text not distorted
  • 2K template runs successfully: denoise 0.33, output around 2K resolution, save prefix z-image-upscaled
  • ImageCompare triple check passed: small text, label edges, material texture
  • Can explain when to switch to the cloud: use Magnific Precise / Nano Banana Pro when VRAM is insufficient

Pitfall reminders ​

❌ Using "creative upscaling" on product images, logo gets redrawn ​

Creative upscaling models reimagine details, which can cause hallucinations and structural drift — product images need fidelity. The official recommendation for product photography is conservative upscaling; Z-Image-Turbo 2K also has a creative component, so keep denoise within 0.35.

❌ Treating upscaling as retouching, expecting it to fix AI artifacts ​

For problems like extra fingers, distortions, or broken images, the official advice is to fix them first with a retouching model before upscaling (or simply regenerate) — upscaling cannot solve structural problems.

❌ Model placed in the wrong directory, not found in the dropdown ​

Put both .safetensors and .pth in ComfyUI/models/upscale_models/; the three Z-Image files each go in their own directories (see the table in step 3). After placing them, refresh or restart ComfyUI for them to appear.

❌ Input image too small, heavy smearing after upscaling ​

The 2K template first normalizes the input to 1 megapixel — when the input is far smaller than this, it amounts to stretching first and then upscaling, and the details cannot be recovered. Try to provide input images around 1 megapixel.

Advanced tips (optional) ​

  • Chained upscaling: if one upscale node isn't enough, chain two (official tips: 2x → 4x), e.g. first RealESRGAN ×4, then a 2x model
  • Generation → upscaling in one pipeline: connect the upscale node directly after the text-to-image output, so images come out in high resolution without manual handoff
  • Cutout → upscale → composite: use the transparent-background PNG cut out in the previous lesson directly as input for conservative upscaling (transparency is preserved), then composite a new background after upscaling — all three steps stay local, your product images never leave home (previous lesson: Product cutout and background replacement)

Lesson summary ​

You've now tried both local routes plus one cloud alternative — for you, ComfyUI product image upscaling is just a matter of picking a model and tuning denoise. ✅ You can now:

  • Do conservative upscaling with the Load Upscale Model + Upscale Image (Using Model) two nodes
  • Load the official Z-Image-Turbo 2K template to add detail up to 2K
  • Control the "drift" risk with the denoise table
  • Switch to cloud-based conservative upscaling when VRAM is insufficient

Key parameter quick reference:

ParameterTemplate defaultNotes
denoise0.330.15–0.25 stays close to the original; 0.25–0.35 adds detail; >0.35 requires a detailed description
steps / cfg5 / 1Z-Image-Turbo distilled version outputs in few steps, cfg fixed at 1
Sampler / schedulerdpmpp_2m_sde / betafixed by the template
Input normalization1 megapixelImageScaleToTotalPixels (lanczos)
Pre-upscaleRealESRGAN ×4 → downscale 0.5net 2x, output at 2K level
Positive promptmasterpiece, 8kswitch to a detailed description at high denoise
Negative promptleft emptytemplate default

Next lesson preview ​

Now that the images are sharp enough, the next lesson puts them "on" a person: Virtual try-on and hand-held product images: Bria and Kling. You will learn:

  • Bria virtual try-on: "dress" your apparel products onto portrait photos
  • Hand-held product image compositing: a model holding your product in a scene shot
  • How to choose the Kling try-on route
  • Usage notes for credit-based billing on cloud nodes (a tip in advance: empty-handed portraits work best, prepare 1–3 photos of the same product from multiple angles)
gaodong.aiAboutPrivacyContact

Independent tutorials, not affiliated with any product mentioned. Product names and trademarks belong to their owners. Check official pages for prices and policies.