# Krea 2 vs Flux.1 Dev Fill vs Qwen Image Edit - Inpainting Pixel Art Test

## Flux.1 Dev Fill

SpriteDX uses Flux.1 Dev Fill model to fill out the following template.

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/87b0c0d9-7912-40ef-b7ad-5f1fec072204.png align="center")

Flux 1. Dev does an awesome job at generating perfectly sized characters for this empty lot.

**Flux.1 Dev Results**:

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/36ea6c96-215a-4ea6-8346-c8c32fc4adc0.png align="center")

All characters come from coherent style and scale.

I wanted to see if there are any alternative to Flux.1 Dev Fill.

## Krea 2

Krea 2 is the closest thing I found.

**Workflow**:

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/0e78dd81-8205-4ada-8529-0eb8ad4efca1.png align="center")

**Krea 2 Configuration**:

|  | Setting |
| --- | --- |
| Model | Krea 2 Raw FP8 |
| Inpainting LoRA | `krea2_inpaint_edit`, Strength **1.0** |
| Text Encoder | Qwen3-VL 4B FP8 · Type `krea2` |
| VAE | Qwen Image VAE |
| Model Patch | Ostris Edit · **KV cache** |
| Sampling | **52 steps / CFG 3.5 / Euler ancestral / simple** |
| Seed | **42**, fixed |

**Krea 2 Results**:

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/6673be64-be47-436f-abfb-faa7dde6b559.png align="center")

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/aa5c2238-2475-450b-818a-20925b36a9a1.png align="center")

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/05e3b387-478f-4d16-b20f-ea95d1a971a2.png align="center")

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/dc14fbb8-7da1-45d3-9cb7-343d0d006744.png align="center")

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/356c2b30-0574-4898-98f1-22a61bdc293d.png align="center")

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/25fc7e70-295d-41fa-9838-37f68bc44f1f.png align="center")

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/ae7f744d-35ab-4dd2-a4aa-f87970e97e09.png align="center")

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/75df077a-32e4-4cde-b9de-4f936450e420.png align="center")

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/0c173299-5a3f-4010-85f8-8ea868b49573.png align="center")

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/ddd7316a-4add-4507-9303-d8a049620944.png align="center")

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/d5e46201-a505-498f-a965-fd8e1e3bcad5.png align="center")

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/3f010cbb-c4d1-4725-a81e-89c4f2dd8e23.png align="center")

**More Results**:

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/e7742dfa-367b-4307-b3b3-81d467a9998a.png align="center")

Included some other prompts in the mix at the end.

## **Qwen Image Edit 2511**

Throwing in a contender.

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/e9d34bff-d61a-4780-863c-d019c7c73e02.png align="center")

**Configuration**:

| Setting | Value |
| --- | --- |
| Model | `qwen_image_edit_2511_fp8mixed.safetensors` |
| Text encoder | `qwen_2.5_vl_7b_fp8_scaled.safetensors` |
| VAE | `qwen_image_vae.safetensors` |
| Steps / CFG | **40 / 4** |
| Sampler / Scheduler | **Euler / simple** |
| Denoise / Flow shift | **1.0 / 3** |
| Seed | **42**, fixed |
| LoRA | None |
| Template | Seven-character sheet, padded to **1456 × 736** |
| Mask | Inverted template alpha → latent noise mask |
| Output | Full sheet + **364 × 364** character crop |

**Qwen Image Edit Results**:

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/4f4d5229-d178-43e9-b681-7aa79266c104.png align="center")

## Qwen Image + Fun Union

There is also option of doing inpainting with Qwen Image. Very similar results.

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/f5e147c5-3340-4017-936f-12ff34291bc1.png align="center")

## Z Image

Don't have time to write about it but style consistency seems superior in Z Image case.

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/380882ed-4b25-4ded-89fa-862c745851b7.png align="center")

But, that's just one sample.

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/60dbb4eb-8e3d-43f4-aa73-c742541f620b.png align="center")

## Discussions

**Analysis**:

*   While Krea 2 generates beautiful pixel arts, there is certain level of style drifts depending on seeds. I think there may be ways to control it better.
    
*   Krea 2 is overall promising alternative if we can contain the style drifts.
    
*   Qwen Image Edit needs character designs are drifting from original designs. This too must be studied further to see if we can control them better. Krea 2 seems to provide better character designs.
    

**Next Steps**:

*   Figure out how to remedy the style drifts.
    
*   Consider replacing Flux [1.Dev](http://1.Dev) Fill with this.
    
*   Also write about Qwen Image Edit 2511 experiments.
    
*   Look into fine-tuning.
    
*   Try Qwen-Image-2512 + Fun ControlNet Union
    
*   Let's also document how zImage does in inpainting.
    
*   look into kimodo flows
    
*   Do CFG and Scheduler variations on various inpainting models.
    
*   Read whitepapers.
    

‒ Sprited Dev 🐛