---
license: other
license_name: qwen-research-license
base_model: Qwen/Qwen-Image-2.1
pipeline_tag: image-to-image
tags:
  - lora
  - outpainting
  - uncrop
  - qwen-image
  - qwen-image-2.1
  - comfyui
  - ai-toolkit
---

# Qwen Image 2.1 Outpaint LoRA

Extends a picture in **any direction**: one side, two sides, a corner, three
sides, all four, or a small picture on a big empty canvas. Pad the picture with
flat gray `#808080`, give the padded canvas to Qwen Image 2.1 as the reference,
and the LoRA fills the gray with a continuation of the scene while keeping
the original picture where it is.

**Ready-made ComfyUI workflow:** [Qwen Image 2.1 Outpaint: auto prompt, any
side, optional LoRA](https://civitai.com/models/2964810/qwen-image-21-outpaint-auto-prompt-any-side-optional-lora) on Civitai.

![padded input | no LoRA | v1 | v2](comparison.jpg)

*Pictures neither LoRA was trained on, rendered in ComfyUI with the INT8 Qwen
Image 2.1 model: 25 steps, CFG 1, seed 42, the same auto-caption prompt in
every column, original pixels pasted back with Stitch Inpaint. Each row:
padded input | no LoRA | v1 | v2. The amber box marks what goes wrong without
the LoRA: visible seams where the original is pasted back, and a duplicated
person.*

## Files

| file | notes |
|---|---|
| `qwen-image-2.1-outpaint.safetensors` | **v1**, step 1500: any extension size, trained at ~1 MP |
| `qwen-image-2.1-outpaint-v2.safetensors` | **v2**, step 2000: everyday extensions at 1-2 MP, people and outfits |
| `checkpoints/qwen21_outpaint_v1_000001250.safetensors` | step 1250, practically identical |
| `checkpoints/qwen21_outpaint_v1_000000500.safetensors` | step 500, already fixes framing; plainer fills |

Both are rank 32, ComfyUI key format (`diffusion_model.transformer_blocks.*`), all 384
tensors load onto the Comfy-Org Qwen Image 2.1 weights.

## v1 or v2

Both fix the same thing (the picture stays in place, so the stitch is clean)
and measure almost the same. v2 is a little better on small and medium
extensions and was trained at 1-2 MP on a people- and fashion-heavy set; v1
saw more extreme zoom-outs.

| unseen pictures | no LoRA | v1 | v2 |
|---|---|---|---|
| subtle crops (10 pictures, ~75 % kept): picture stays in place | 16.4 dB | 34.6 dB | 34.7 dB |
| subtle crops: fill error vs the real photo (lower is better) | 16.3 | 10.5 | 10.2 |
| big crops (8 pictures, 36-50 % kept): picture stays in place | 25.5 dB | 33.9 dB | 34.0 dB |
| big crops: fill error | 29.8 | 27.0 | 27.2 |

## Prompt

The instruction is the trigger. Put it first:

```
Outpaint the image: replace the solid gray areas with a seamless continuation of the scene, keeping the existing picture unchanged.
```

Optionally follow it with `Scene: <description>`. The training captions
described the whole picture, so a description of the source (what is visible)
works well; in ComfyUI, **Text Generate** with the Qwen3-VL 8B encoder you
already load for Qwen 2.1 can write it automatically from the picture.

## ComfyUI

The quickest start is the [ready-made workflow on Civitai](https://civitai.com/models/2964810/qwen-image-21-outpaint-auto-prompt-any-side-optional-lora). To build
it yourself:

1. Pad the picture with flat gray `#808080` (AusBoss **Load Image + Pad**:
   fill `color`, `#808080`, canvas multiple 32, target 1.0 MP for v1, 1-2 MP for v2).
   For a tilted picture, use **Image Crop + Rotate + Pad** with `feather` 0 instead
   (see [Tilted pictures](#tilted-pictures)).
2. **Text Encode Qwen Image 2.1**: the padded canvas as `image_1`,
   `resolution` 0 (reference and output share the canvas size), plus the prompt.
3. **KSampler** on the encoder's `latent` output: 25 steps, CFG 1,
   `euler` / `simple`, denoise 1.
4. **VAE Decode** -> **Split Image with Alpha** (the Qwen 2.1 VAE decodes RGBA)
   -> AusBoss **Stitch Inpaint** to paste the exact original pixels back over a
   32 px feather.

Do **not** pin the known area with *Set Latent Noise Mask*: on Qwen 2.1 that
draws a visible rectangle at the seam. The LoRA keeps the picture in place on
its own.

The Civitai workflow's LoRA Loader skips a missing file instead of stopping, so
without the download it quietly runs plain Qwen 2.1. If your results show a
lighter box or doubled edges where the original was pasted back, check that
the LoRA row is loaded.

## What v1 fixes (12 held-out pictures, ComfyUI renders)

| | base Qwen 2.1 | + LoRA 500 | + LoRA 1250 | + LoRA 1500 |
|---|---|---|---|---|
| gray left unfilled | 9.1 % | 1.1 % | 1.0 % | 0.9 % |
| kept-area PSNR (picture stays in place) | 23.6 dB | 33.6 dB | 33.6 dB | 33.5 dB |
| mean colour step at the seam | +2.2 | -0.8 | -0.7 | -0.6 |

Without the LoRA, Qwen 2.1 often reframes or rescales the picture (a fisheye
kitchen shrank inside its frame, a street scene was recomposed) or returns a
gray frame untouched. With it, the known picture stays pixel-registered, which
is what makes a clean stitch possible. (The "gray" metric counts any flat
mid-gray, so gray pavement or sky can read as a few percent.)

## Tilted pictures

Straightening or tilting a picture before outpainting leaves gray wedges in the
corners instead of straight borders. Neither LoRA was trained on diagonal
borders, and both handle them.

![tilted input | no LoRA | v1 | v2](rotation.jpg)

*36 ComfyUI renders: a sports photo, a flat lawn and the pier above, each
rotated 17° with AusBoss Image Crop + Rotate + Pad, 2 seeds, `feather` 24 and 0,
the same prompt and seed in every column, original pasted back with Stitch
Inpaint.*

| rotated 17° | no LoRA | v1 | v2 |
|---|---|---|---|
| renders where the picture did not stay in place near the tilted edges (> 2 px) | 12 / 12 | 0 / 12 | 0 / 12 |
| largest shift (sports photo, pier) | 95 px | 0 px | 0 px |

Without the LoRA, Qwen 2.1 moves the scene while it fills the wedges, so the
pasted-back original no longer lines up: a lighter tilted box, cut or doubled
edges and ghosted objects along the tilt. With either LoRA the picture stays
exactly in place.

Set `feather` to **0** on Image Crop + Rotate + Pad (AusBoss nodes 2.2.0). Its
feather also fades the picture itself into the gray fill (Load Image + Pad only
feathers the mask). The model paints that ramp as a darker band along the tilt,
and Stitch Inpaint's color match reads the faded edge and pulls flat areas
grayer. Stitch Inpaint still blends the paste-back over 32 px at feather 0.

## Training

### v1

- [ostris/ai-toolkit](https://github.com/ostris/ai-toolkit), `arch: qwen_image_2`,
  Comfy-Org INT8 convrot base (the same weights ComfyUI runs), text encoder
  embeddings cached with the reference image.
- 924 pairs from 231 pictures, 4 layouts each: all-sides frame, one side,
  opposite sides, corner, three sides, small window (12-30 % kept). Target =
  the picture at <= 1 MP on the /32 grid (never upscaled), source = same canvas
  with everything outside the kept rectangle painted `#808080`.
- Captions: the instruction (70 %) or one of five paraphrases (30 %), and for
  75 % of pairs `Scene:` + a Qwen3-VL-8B description of the whole picture.
  Caption dropout 0 (in this trainer a dropped caption also drops the reference).
- Rank 32 / alpha 32, AdamW8bit, LR 1e-4 constant, batch 1, `shift` timesteps,
  1500 steps. Converged by about step 1250; step 500 already works.
- Pictures: a hand-picked set plus openly licensed images: Flickr photos from
  [CommonCatalog CC-BY](https://huggingface.co/datasets/common-canvas/commoncatalog-cc-by),
  museum art from [PD12M](https://huggingface.co/datasets/Spawning/PD12M)
  (CC0 / public domain), anime from
  [anime-with-caption-cc0](https://huggingface.co/datasets/alfredplpl/anime-with-caption-cc0).

### v2

- 68 hand-picked pictures: 47 of people and outfits (Unsplash photos under the
  [Unsplash License](https://unsplash.com/license) via
  [unsplash-lite](https://huggingface.co/datasets/1aurent/unsplash-lite), plus
  the author's own), 21 of scenes, interiors, anime and paintings. One random
  layout per picture, 68 pairs.
- Smaller extensions than v1: the kept picture covers 45-93 % of the canvas
  (median ~74 %), about one layout in seven bolder; no small-window zoom-outs.
- Targets at 1-2 MP (never upscaled), ai-toolkit `resolution: 1408`.
- Same captions recipe, rank 32 / alpha 32, AdamW8bit, LR 1e-4, 2000 steps.
  Works from step 500; step 2000 had the lowest fill error.

## Limits

- Very large extensions (the kept picture under ~15 % of the canvas) invent a
  lot; results vary more by seed.
- v1 was trained at about 1 MP, v2 at 1-2 MP; much larger canvases are untested.
- Text in the new area is plausible, not legible.
- On shallow-focus photos, tilt wedges come out a little softer than an
  in-focus subject: the model continues the blur it sees. Qwen 2.1 without the
  LoRA, Krea 2 and FLUX.2 Klein did the same in the same test.

## License

A LoRA for Qwen Image 2.1, which is released under the Qwen Research License;
use of the base model, and of this LoRA with it, follows that license.