#### NOTE
This is a **community plugin**, an external project maintained by its respective author.
Community plugins are not part of FiftyOne core and may change independently.
Please review each plugin’s documentation and license before use.

<a href="https://github.com/harpreetsahota204/qwen_image_edit" target="_blank">![GitHub Repo](https://img.shields.io/badge/GitHub-Repository-black?logo=github)</a>

# Qwen Image Edit Panel

![image](https://raw.githubusercontent.com/harpreetsahota204/qwen_image_edit/main/qwen_image_edit_panel.gif)

A [FiftyOne](https://docs.voxel51.com) plugin that brings chat-based image editing directly into the sample modal. Write a prompt, see all generated results inline, iterate, and save your favourite edits back to your dataset — all without leaving the app.

Inference runs **entirely on your local GPU** using the [Qwen Image Edit 2511](https://huggingface.co/drbaph/Qwen-Image-Edit-2511-FP8) model in FP8 quantized form. No API keys or external services required.

## Installation

Download the plugin:

```bash
fiftyone plugins download https://github.com/harpreetsahota204/qwen_image_edit --overwrite
```

Set the operator timeout **before** launching FiftyOne. The default 600-second timeout is not enough for a cold-start model load (up to ~10 min on first run) or a long multi-step inference call (1–5 min per edit depending on steps and image count):

```bash
export FIFTYONE_OPERATOR_TIMEOUT=1800   # 30 minutes
```

Then launch the app:

```python
import fiftyone as fo
import os

os.environ["FIFTYONE_OPERATOR_TIMEOUT"] = "1800"

dataset = fo.load_dataset("your-dataset")
fo.launch_app(dataset)
```

Or from the shell:

```bash
export FIFTYONE_OPERATOR_TIMEOUT=1800
fiftyone app launch
```

> **Important:** `FIFTYONE_OPERATOR_TIMEOUT` must be set in the same environment that runs the FiftyOne server, not just the Python session that opens the browser. Per-function timeout decorators from `fiftyone.operators.decorators` use Python’s `signal` module and do not work in FiftyOne’s operator worker threads — the environment variable is the only supported mechanism.

---

## Requirements

| Requirement   | Detail                                 |
|---------------|----------------------------------------|
| FiftyOne      |                                        |
| PyTorch       | CUDA build (`torch`)                   |
| Diffusers     | `diffusers`                            |
| Accelerate    | `accelerate`                           |
| Pillow        | `Pillow`                               |
| GPU           | CUDA-capable, ≥ 80 GB VRAM recommended |

Install Python dependencies:

```bash
pip install torch diffusers accelerate Pillow fiftyone
```

---

## What it does

Open any sample in the FiftyOne modal and the **Qwen Image Edit** panel appears alongside it. From there you can:

- **Prompt-driven edits** — describe the change you want in plain text and the panel runs the Qwen FP8 pipeline locally on your GPU to apply it.
- **Multiple images per prompt** — set *Images per prompt* to generate several variations in one call. Each is shown as a separate card so you can compare them side by side in the panel.
- **Iterative refinement** — each edit turn is added to the panel’s history. Click the \*\*\*\* (set as source) button on any image — from any turn — to use it as the starting point for your next edit.
- **Advanced controls** — optionally supply a negative prompt, override inference steps, adjust *True CFG scale* (Qwen’s classifier-free guidance strength), and fix a random *Seed* for reproducibility.
- **Per-image actions** — every generated image has its own **Save**, **Trash**, and \*\* Set as source\*\* buttons. Save any subset you like; trash the ones you don’t want to keep.
- **Save to dataset** — clicking Save copies the image into your dataset’s media directory and registers it as a new group slice (`edit_1`, `edit_2`, …). If the dataset is flat it is automatically converted to a grouped dataset on first save.
- **Label copying** — when saving, optionally copy any label fields (detections, classifications, segmentations, etc.) from the source sample onto each new edited slice.
- **Session memory** — your edit history for each sample is preserved in the browser session, so switching between samples and coming back doesn’t lose your work.
- **Live status** — during the first run the panel shows real-time download and GPU-loading status. During inference it shows step-level progress (e.g. `Step 11/40 (28%)`) with an elapsed-time counter.

---

## Model

|               |                                                                                                                                       |
|---------------|---------------------------------------------------------------------------------------------------------------------------------------|
| **Repo**      | [`Qwen/Qwen-Image-Edit-2511`](https://huggingface.co/Qwen/Qwen-Image-Edit-2511)                                                       |
| **Weights**   | [`drbaph/Qwen-Image-Edit-2511-FP8`](https://huggingface.co/drbaph/Qwen-Image-Edit-2511-FP8) — FP8 E4M3FN quantized transformer, ~4 GB |
| **Precision** | `torch.bfloat16` compute, `torch.float8_e4m3fn` storage                                                                               |
| **GPU**       | CUDA-capable GPU with at least 80 GB VRAM recommended                                                                                 |

The transformer weights are downloaded on the **first edit only** and cached by HuggingFace Hub (`~/.cache/huggingface/`). All subsequent runs load directly from cache. The panel shows download and loading status live while you wait.

---

## First-run note

The first time you submit a prompt the panel will:

1. Download the FP8 transformer checkpoint (~20.4 GB) from HuggingFace Hub — this is a one-time step and may take several minutes depending on your connection.
2. Load the full pipeline onto your GPU — this can take several minutes.
3. Begin denoising — progress is shown step by step (e.g. `Step 1/40 (2%) … Step 40/40 (100%)`).

All subsequent edits in the same session skip steps 1 and 2 and go straight to inference.

---

## Getting Started

If you want to quickly test the plugin, you can use the FiftyOne quickstart dataset:

```python
import fiftyone as fo
import fiftyone.zoo as foz

dataset = foz.load_zoo_dataset("quickstart")
session = fo.launch_app(dataset)
```

Alternatively, if you have your own dataset and want to test the plugin with your data, see the [FiftyOne documentation on importing datasets](https://docs.voxel51.com/user_guide/import_datasets.html) to learn how to load your data into FiftyOne.

## Usage

1. Open a dataset in the FiftyOne app and click into any sample.
2. Open the **Qwen Image Edit** panel from the modal panel selector.
3.  *(Optional)* Click \*\* Advanced\*\* to set a negative prompt, inference steps, True CFG scale, seed, or images-per-prompt count.
4. Type a prompt describing your edit and press **Enter** or click \*\*\*\*.
5. The panel shows live status during download/loading and step-by-step progress during inference.
6. When results appear, each generated image has three buttons:
   - \*\*\*\* — set this image as the source for your next edit.
   - \*\* Trash\*\* — delete the temp file and remove the card.
   - \*\* Save\*\* — persist the image as a new `edit_N` group slice on the dataset.
7. Click \*\*\*\* on any image (including the original) to branch from it, then type a new prompt and repeat.

### Advanced parameters

| Parameter         | Default   | Description                                                                                 |
|-------------------|-----------|---------------------------------------------------------------------------------------------|
| Negative prompt   | *(empty)* | Concepts to suppress. Sent as a single space when blank (Qwen requires a non-empty string). |
| Steps             | 40        | Number of denoising steps. More steps = higher quality, slower inference.                   |
| True CFG scale    | 4.0       | Classifier-free guidance strength specific to Qwen.                                         |
| Seed              | 51        | Random seed. Fix this to get reproducible results across calls.                             |
| Images per prompt | 1         | Number of images to generate per edit. All are shown as separate cards.                     |

---

## Saved sample fields

Each saved edit slice carries the following fields in addition to the image:

| Field                      | Type   | Description                                            |
|----------------------------|--------|--------------------------------------------------------|
| `edit_prompt`              | String | The prompt used for this edit.                         |
| `edit_model`               | String | Model identifier.                                      |
| `edit_negative_prompt`     | String | Negative prompt (if any).                              |
| `edit_num_inference_steps` | Int    | Steps used.                                            |
| `edit_true_cfg_scale`      | Float  | True CFG scale used.                                   |
| `edit_seed`                | Int    | Seed used.                                             |
| `generation_time`          | Float  | Wall-clock seconds the pipeline call took.             |
| `edit_history`             | List   | Full turn history (prompt + filepath) for the session. |
| `tags`                     | List   | Always includes `"edited"`.                            |