# Run Qwen-Image 2.1 locally and inspect your first PNG

> Install local ComfyUI, load the pinned base Qwen-Image 2.1 workflow and weights, then generate, reopen and check one saved PNG from your own prompt.

By BIG CHANGE Editorial

Published: 2026-10-21T21:08:38.225Z
Updated: 2026-10-21T21:08:38.225Z
Canonical: https://bigchange.ai/blog/qwen-image-2-1-local-comfyui-first-png

![Conceptual charcoal illustration of a generic orange pear image on an unbranded desktop monitor, with a magnifying glass resting on the table in front.](https://bigchange.ai/api/media/file/qwen-local-png-inspection-hero-v1.png)
AI-generated conceptual illustration by BIG CHANGE.

Illustration: The pear on the monitor is AI-generated conceptual artwork by BIG CHANGE. It is neither Qwen-Image 2.1 output nor a BIG CHANGE test result.

ComfyUI now has a native Qwen-Image 2.1 text-to-image workflow. This guide takes you from a local installation to one image you can actually open and inspect. You need a computer that can run the selected model, room for the weights and a prompt you are permitted to use. The steps follow ComfyUI's [current manual](https://docs.comfy.org/tutorials/image/qwen/qwen-image-2-1) and a [pinned base template](https://github.com/Comfy-Org/workflow_templates/blob/7d8ed32f0bf6b8d52129976535422d6f518fe463/templates/image_qwen_image_2_1_t2i.json); BIG CHANGE has not run the model or measured its hardware requirements.

## The big change

- **What changed:** ComfyUI supplies a native local workflow for Qwen-Image 2.1. Its base text-to-image graph connects the model, encoder, resolution control and PNG save step, giving a local user a reproducible starting point.
- **Why it matters:** A user doing research or evaluation can keep a permitted prompt and image on their own machine and inspect the saved result. The workflow still needs large downloads and enough working memory; a template that opens is only the start.
- **What to watch:** The installed ComfyUI build must contain the graph's nodes, and the selected weights and research license must fit the intended use. Judge the first run by the actual PNG's dimensions and visible details, not the preview or a model capability claim.

## Before you download

Choose a self-authored image brief whose details you can check visually. Record its exact wording, the requested objects and composition, and what would make the result unusable. Use only prompts and any assets you have rights to use. This base route takes a text prompt; it does not need an input portrait or an image-edit graph.

The three selected [Comfy-Org model files](https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/a8342359c61979b06a8dafb1a47991fe1f30e994) total 17,283,091,112 bytes in the checked repository metadata. Leave more disk space for downloads, the installation, working files and PNG output. That figure is **storage**, not a VRAM or RAM requirement. Check the machine's available memory and ComfyUI's [system and installation guidance](https://docs.comfy.org/installation/desktop/overview) before committing to the downloads. No GPU type, speed or successful run is guaranteed by the file sizes.

Read the model's [Qwen Research License](https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE) before use. Its grant covers noncommercial research or evaluation of the materials; commercial use requires a separate license. Check the current terms and rights for your intended inputs and outputs. Running the files locally avoids a hosted per-image API charge on this route, but uses your bandwidth, storage, electricity and hardware.

## 1. Install or update local ComfyUI

For a fresh install, [Comfy Desktop](https://docs.comfy.org/installation/desktop/overview) can create and manage a local ComfyUI instance on its supported platforms. Follow its platform installation steps and keep track of the instance and engine version. For a [manual installation](https://docs.comfy.org/installation/manual_install), create an isolated Python environment, clone ComfyUI and install the GPU dependencies specified for your actual device. Activate that same environment and open a terminal in the cloned `ComfyUI` directory. Install ComfyUI's requirements there:

```shell
pip install -r requirements.txt
```

From the same directory and environment, start the local application:

```shell
python main.py
```

Use the current platform instructions for the GPU dependency commands rather than assuming another device's packages apply to yours.

If ComfyUI is already installed, use the [update instructions for that installation type](https://docs.comfy.org/installation/update_comfyui). Desktop's Manage panel has an engine Update tab with Stable and Latest on GitHub channels; the Desktop app's own update is separate. For a manual Git installation, open a terminal in its existing `ComfyUI` directory and activate that installation's Python environment. Update the code and requirements in that order:

```shell
git pull
pip install -r requirements.txt
```

Restart ComfyUI and record the version or commit you actually ran. Stable builds can trail new native nodes. If the workflow reports missing nodes, inspect startup import errors and the installed build. Use a documented current channel compatible with the graph or wait for stable support.

## 2. Load the base text-to-image graph

On the [pinned `image_qwen_image_2_1_t2i.json` page](https://github.com/Comfy-Org/workflow_templates/blob/7d8ed32f0bf6b8d52129976535422d6f518fe463/templates/image_qwen_image_2_1_t2i.json), use GitHub's **Download raw file** control and keep the `.json` filename. Drag that downloaded JSON onto the local ComfyUI canvas to load it. ComfyUI's workflow examples document dragging a downloaded JSON into the canvas; its [Workflows guide](https://docs.comfy.org/interface/workflows) also covers opening saved graphs. The Template Library lists Qwen-Image-2.1 text to image if you prefer it, but compare its graph and settings with the pinned version and record differences. Select the **base text-to-image** graph, rather than an edit or Turbo template.

The pinned top level has `ResolutionSelector` node 13 feeding width and height into the Qwen Image 2.1 subgraph, followed by `SaveImageAdvanced` node 461. The selector starts at `1:1 (Square)`, `1` megapixel and a multiple of `8`. That targets roughly 1024 × 1024 pixels. The linked selector controls the run; width and height values visible inside the subgraph do not override those connections. Leave this first run at the linked square target and record the dimensions the selector shows.

## 3. Place and select the three base files

From the pinned [Comfy-Org Qwen-Image-2.1 model tree](https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/a8342359c61979b06a8dafb1a47991fe1f30e994), download these exact files into the selected local instance's model folders:

| Loader input | File | Directory under `ComfyUI/models/` |
| --- | --- | --- |
| Base diffusion model | `qwen_image_2.1_int8_convrot.safetensors` | `diffusion_models/` |
| Qwen3-VL text encoder | `qwen3vl_8b_int8_convrot.safetensors` | `text_encoders/` |
| Qwen-Image VAE | `qwen_image_2.1_vae_bf16.safetensors` | `vae/` |

Check each filename and the repository revision before download, then confirm the completed file's identity and size. Refresh model discovery or restart ComfyUI. On the canvas, locate the large Qwen Image 2.1 subgraph between `ResolutionSelector` and `SaveImageAdvanced`. Its exposed model selectors are `unet_name`, `clip_name` and `vae_name`; choose the three files in the table there. If those controls are hidden in your layout, [ComfyUI's subgraph guide](https://docs.comfy.org/interface/features/subgraph) says to double-click an empty area inside the subgraph, avoiding its widgets, or use its edit button to enter it. The guide also documents the **Edit Subgraph Widgets** parameters panel on supported builds. Return to the parent graph with the navigation bar or **Esc**. The optional `PE_model` selector may still name `qwen3.5_9b_qwen_image_2.1_pe_t2i.int8_convrot.safetensors`. It is the separate prompt-enhancement encoder. With `refine_prompt` **false**, as in this template, that branch does not run and this fourth file is not needed for the first base image. If you later turn the switch on, obtain the matching file and revisit the changed prompt path.

## 4. Set the prompt and run once

In that same Qwen Image 2.1 subgraph, replace its exposed `prompt` field's demonstration text with **your actual permitted brief**. Keep a copy outside the graph, including the concrete details you plan to inspect. Record the visible `seed` and any value you change. Leave `refine_prompt` false so the image model receives the typed prompt without the optional rewrite. Use the subgraph edit controls described above if a field is hidden.

For this first run, keep the pinned base values: `steps` **25**, `cfg` **1**, sampler **euler**, scheduler **simple**, and the square `ResolutionSelector` at **1 megapixel**. At `cfg` 1 the negative-conditioning pass is skipped, so changing the negative prompt does not help this run. Qwen-Image 2.1 Turbo is a different checkpoint with an eight-step `ManualSigmas` schedule; do not bring that schedule into this base graph.

Back on the parent graph, click **Run** once, or press **Ctrl+Enter** on Windows/Linux or **Cmd+Enter** on macOS. ComfyUI's [workflow execution guide](https://docs.comfy.org/tutorials/api-nodes/luma/luma-text-to-image) documents that control and shortcut. Watch for an execution error and make sure the graph reaches `SaveImageAdvanced`, which the pinned template sets to `png`, `8-bit`, `sRGB` with filename prefix `Qwen_image_2.1`. A loaded graph or an on-screen preview does not establish that a file was saved.

## 5. Reopen the saved PNG and decide what worked

Look first in the local instance's default `ComfyUI/output/` folder for the file produced by `SaveImageAdvanced`. If your Desktop instance or startup configuration uses another output directory, inspect that instance's configured output location. Use the actual filename shown for the run, since ComfyUI can add numbering. Open that PNG in an image viewer. Confirm it decodes and record its pixel width and height. Compare them with the linked resolution target. Then check the objects, any requested text, placement and other visible details against **the brief you actually entered**. Note missing details and artifacts. Do not substitute the template's sample prompt or claim its sample image as your result.

Keep a short run record: ComfyUI version or commit and install route; template revision; three file names or revisions; exact prompt; seed; selector size; steps, `cfg`, sampler and scheduler; saved filename and measured dimensions; and what you saw. The task is complete when your real PNG opens and that comparison is recorded, even if the image fails your quality criterion. If you try again, change one prompt detail or setting, keep the seed fixed when comparing settings, and inspect the new saved file separately.

## When the first run fails

- **A node is missing:** Check the ComfyUI build and startup import log, then update the engine and dependencies through the documented route for your installation. The current manual warns that a stable build may lack a node available in a newer build.
- **A loader cannot find a model:** Check the selected instance, exact folder and filename, and the pinned model revision. Refresh or restart model discovery. Verify that the diffusion file is the base `qwen_image_2.1_int8_convrot` file and that its encoder and VAE match, before diagnosing the sampler.
- **Memory runs out or the run stalls:** Record the device, available RAM and VRAM, resolution and exact error. Try a lower megapixel target or a documented memory mode supported by your build, then compare the actual output dimensions. The download size cannot diagnose runtime memory fit.
- **The image looks wrong or has unexpected dimensions:** Check the selector's width and height links, the saved seed and the prompt actually sent to the graph. Reopen the PNG rather than relying on the preview. The [manual](https://docs.comfy.org/tutorials/image/qwen/qwen-image-2-1) also describes fine texture artifacts from the stock VAE; changing sample steps alone does not repair a decoder artifact.

This is a documentation-led route. BIG CHANGE did not install ComfyUI, download the weights, generate a PNG, inspect output quality, time a run or establish a minimum compatible GPU. The manual describes capabilities such as native 2K output, typography and alpha; this first square 1-megapixel run tests only what your own saved file shows.

## Sources

- [ComfyUI Qwen-Image-2.1 native workflow manual](https://docs.comfy.org/tutorials/image/qwen/qwen-image-2-1) — Base versus edit/Turbo workflow, model paths, sampling, resolution, enhancement and troubleshooting; undated current manual.
- [Pinned Comfy-Org Qwen-Image 2.1 base text-to-image template](https://github.com/Comfy-Org/workflow_templates/blob/7d8ed32f0bf6b8d52129976535422d6f518fe463/templates/image_qwen_image_2_1_t2i.json) — Commit-specific linked ResolutionSelector, filenames, base defaults and PNG SaveImageAdvanced; current library may differ.
- [Pinned Comfy-Org Qwen-Image 2.1 model tree](https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/a8342359c61979b06a8dafb1a47991fe1f30e994) — Three selected file identities and metadata sum 17,283,091,112 bytes; storage only, no runtime-memory measurement.
- [Comfy Desktop overview](https://docs.comfy.org/installation/desktop/overview) — Local Desktop instance installation, platform scope and separate ComfyUI engine management.
- [ComfyUI manual installation](https://docs.comfy.org/installation/manual_install) — Isolated environment, repository, device-specific dependencies, startup and manual update.
- [How to update ComfyUI](https://docs.comfy.org/installation/update_comfyui) — Documented Desktop, portable and manual engine update channels; stable build limitations.
- [Qwen-Image 2.1 Research License](https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE) — Noncommercial research/evaluation grant for model materials; separate commercial license needed.
- [ComfyUI subgraph guide](https://docs.comfy.org/interface/features/subgraph) — Documented edit button, empty-area double-click, parameters panel limits and parent-graph navigation; no UI run here.
- [ComfyUI official JSON workflow import example](https://docs.comfy.org/tutorials/api-nodes/rodin/model-generation) — Download JSON and drag into ComfyUI; article uses this generic local graph import control only.
- [ComfyUI Workflows guide](https://docs.comfy.org/interface/workflows) — Saved workflow panel and opening graphs; current UI layout may vary.
- [ComfyUI documented Run and output example](https://docs.comfy.org/tutorials/api-nodes/luma/luma-text-to-image) — Run button, Ctrl/Cmd+Enter and default ComfyUI/output folder; configured instance can use another location.