Skip to content
Easy AI How-To

IntermediateDGX SparkImages

Redact personal info from phone screenshots with ComfyUI and Qwen Image Edit on DGX Spark

I'll be honest: I didn't buy my DGX Spark to save money. I bought it to learn by doing.

I bought it last year, when it cost about $4,000. That was roughly six months earlier than I'd planned. Memory prices were starting to bite, and I'd heard the price was about to rise to $4,700. This week the same machine is going for about $7,000, and no, I'm not ready to sell.

I never expected it to save me money on AI subscriptions. If anything, I expected to spend more. What I wanted was to stop reading about AI engineering and start doing it, on hardware I control, where I could break things and learn from them.

The full story is too big for one sitting, so here's the first thing I did. I'd always wanted a simple tool that takes screenshots from my phone and scrubs the personal info: names, phone numbers, emails. One of my first proofs of concept on the Spark used ComfyUI and Qwen Image Edit. It worked perfectly on the first try, and it kept working as I iterated. I was blown away.

My original proof of concept used an earlier Qwen Image Edit release; this recipe uses the newest one, Qwen-Image-Edit-2511. Here's the recipe, so you can do it too. — Greg Alton

More DGX Spark recipes are coming. Follow the DGX Spark section for the next one.

Time
About 60 minutes
Checked
Tools
NVIDIA DGX Spark ComfyUI playbook, ComfyUI Qwen-Image-Edit-2511 tutorial, Qwen-Image-Edit-2511 on Hugging Face

If a button name on the screen does not match this guide, follow the action. These products rename buttons often.

Steps

  1. Install ComfyUI with NVIDIA's DGX Spark playbook

    DGX Spark is an Arm64 (aarch64) machine with a Grace Blackwell chip and 128 GB of unified memory, so generic x86 install guides can trip you up. NVIDIA's official ComfyUI playbook handles it. The Image Gen Quick Start path makes a Python virtual environment (an isolated folder of packages), installs CUDA-enabled PyTorch for the Spark, clones ComfyUI, and installs its requirements. Open the playbook and follow its current steps, or run its automated setup script from a folder in your home directory:

    curl -fsSL "https://raw.githubusercontent.com/NVIDIA/dgx-spark-playbooks/refs/heads/main/nvidia/playbook-comfyui/assets/setup.sh" | bash

    The script also downloads a starter image model (about 20 GB) you can use to confirm everything works. Prefer containers? The playbook's Video Gen Workflow tab has a Docker build on NVIDIA's PyTorch image instead.

  2. Launch it and open the UI

    From the same folder, start the server with the playbook's launch script (or python main.py --listen 0.0.0.0 inside the ComfyUI folder with the venv active). Check it's up:

    curl -I http://localhost:8188

    You should see HTTP/1.1 200 OK. Then open http://<SPARK_IP>:8188 from your laptop's browser. --listen 0.0.0.0 makes ComfyUI reachable from your whole network, so only do this on a network you trust.

  3. Make sure ComfyUI is recent enough

    Qwen-Image-Edit-2511 needs a recent ComfyUI with its native Qwen nodes. If the Templates menu has no Qwen Image Edit 2511 entry, or nodes show up red when you load the workflow, update ComfyUI (in the ComfyUI folder: git pull, then pip install -r requirements.txt with the venv active) and restart.

  4. Download the Qwen-Image-Edit model files

    This recipe uses Qwen-Image-Edit-2511, the current Qwen image-editing release, repackaged for ComfyUI by the ComfyUI team. Put each file in the matching folder under ComfyUI/models/:

    ComfyUI/models/
├── diffusion_models/  qwen_image_edit_2511_bf16.safetensors
├── text_encoders/     qwen_2.5_vl_7b_fp8_scaled.safetensors
├── vae/               qwen_image_vae.safetensors
└── loras/             Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors  (optional, faster)

    Download links for each file are in the ComfyUI tutorial linked above. The bf16 diffusion model is large (it's a 20-billion-parameter model), and the Spark's 128 GB of unified memory is what lets you run it at full precision instead of a smaller cut-down version.

  5. Load the image-edit workflow

    In ComfyUI, open Templates, find Qwen Image Edit 2511, and load it (or drag in the workflow from the ComfyUI tutorial). The graph looks roughly like this:

    Load Image (screenshot) ─┐
                         ├─▶ Text Encode (Qwen edit, your prompt) ─┐
Load CLIP (Qwen 2.5 VL) ─┘                                         │
Load Diffusion Model (2511) ─▶ (optional) Lightning LoRA ─▶ KSampler ─▶ VAE Decode ─▶ Save Image
Load VAE ─────────────────────────────────────────────────────────────┘

    Check that each loader node points to the file you downloaded. If you use the Lightning LoRA, keep the template's low step count; without it, use the template's normal steps.

  6. Write the redaction prompt

    Qwen-Image-Edit follows plain-language instructions and is unusually good with text in images, which is exactly what screenshots are. Be specific about what to change and what to keep.

    Prompt to paste

    Redact all personal information in this phone screenshot. Cover every person's name, phone number, email address, street address, and profile photo with solid black rectangles. Keep everything else exactly the same: layout, app interface, icons, timestamps, message bubbles, and all other text. Do not add anything new.

  7. Try a fake-placeholder version

    Black bars are the safest. For tutorials or demos where you want the screenshot to still look natural, ask for realistic placeholders instead, and use obviously fake values:

    Replace every person's name with "Alex Sample", every phone number
with "(555) 010-0000", and every email with "user@example.com".
Match the original font, size, and color. Keep the layout and all
other text identical.

  8. Run, inspect, iterate

    Click Run (or Queue). Compare the output with the original side by side. If something slipped through, name it in the prompt ("also cover the contact name in the top bar") and run again. Fix the seed in the sampler when you like a result so later tweaks stay comparable.

  9. Batch a folder of screenshots

    Once one prompt works, reuse it. The simplest way: put screenshots in ComfyUI/input/, switch the Load Image node to each file, and queue several runs. For hands-off batches, save the workflow in API format and post it to ComfyUI's /prompt endpoint from a short script, changing the image filename each time. Results land in ComfyUI/output/.

Tips

  • Always check the output by eye before you share it. Image models can miss small text, partly hidden numbers, or names inside notifications.

  • Black boxes are safer than blur. Light blur on text can sometimes be reversed, and placeholders can look like real data if you aren't careful.

  • Keep originals local and delete them when you're done. Running on your own DGX Spark means the screenshots never leave your machine, which is the whole point.

  • Edited images can change pixels you didn't ask about. For anything legal or official, use a dedicated redaction tool and keep a record of what you removed.

Common questions

Why not just use a cloud AI app?

You could, but you'd be uploading the very personal data you're trying to hide. Running locally keeps it on your own hardware.

I heard about 'qwen-image-edit-3702'. Is that a different model?

We couldn't find a release by that name. The Qwen-Image-Edit releases are named by date (2509, 2511, and so on). This recipe uses 2511, the newest Qwen-Image-Edit release on Qwen's Hugging Face page at the time of writing; if a newer one appears in the ComfyUI templates, it should drop into the same workflow.

Does this work without a DGX Spark?

Yes, ComfyUI runs on many NVIDIA GPUs. Smaller cards need the fp8 or int8 versions of the model, which use less memory. The Spark's unified memory just makes the full-size model comfortable.

More guides

Occasional notes

Hear when a new guide is up

A short note when a guide is added. No account, and no emails until this form is connected to a real list.

Placeholder. This form is not connected to an email service. Nothing you type is saved or sent.