Krea2 Super-Resolve
A beta LoRA for Krea 2 image super-resolve, trained on 1,000 fuzzed images from the Aesthetic-4K dataset.
What does it do?
Super-resolution in AI training is the process of teaching a neural network to take a low-resolution image and accurately generate a sharper, higher-resolution version with realistic details. This will attempt to restore images to their original or intended quality, including people, landscapes, art, and brushstrokes.
Model
| Item | Details |
|---|---|
| Weights | Krea2_Super-Resolve-beta.safetensors |
| Model type | LoRA adapter |
| Base model in the training config | Krea 2 Raw |
| Task | Image super-resolve using image-edit training |
| Status | Beta |
This repository provides LoRA weights, not a standalone model. A compatible Krea 2 base model and image-edit inference workflow are needed to use the adapter.
Examples
Source images are shown on the left, with super-resolved results on the right.
Examples 1-5 are examples from low-resolution photographs. Examples 6-10 are outputs from synthetcially fuzzed inputs.
| Source | Super-Resolved |
|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Training
The model was created using conradlocke/krea2-identity-edit and lbouaraba/krea2edit-trainer. The krea2edit-trainer script was adapted to run with Musubi Tuner for this training run.
Dataset and Run
| Setting | Value |
|---|---|
| Dataset | zhang0jhon/Aesthetic-4K |
| Training images | 1,000 fuzzed images from the dataset |
| Epochs completed | 5 |
| Training steps completed | 5,000 |
| GPU | NVIDIA GeForce RTX 4090 |
| Optimizer | Adafactor |
Configuration Details
The following were used with Musubi Tuner:
| Setting | Configured value |
|---|---|
| Training mode / task | LoRA Training / Image Edit |
| LoRA module | networks.lora_krea2 |
| LoRA rank / alpha | 128 / 128 |
| LoRA dropout | 0 |
| Learning rate | 3e-5 |
| LR scheduler | Constant, no warmup |
| Adafactor arguments | scale_parameter=False, relative_step=False, warmup_init=False, weight_decay=0.01 |
| Dataset resolve | 768 x 768, aspect-ratio bucketing enabled |
| Batch size / gradient accumulation | 1 / 1 |
| Mixed precision / save precision | BF16 / BF16 |
| Base-model quantization | FP8 enabled, scaled FP8 enabled |
| Gradient checkpointing | Enabled |
| Blocks to swap | 26 |
| Latent / text-encoder output caching | Both enabled |
| Edit text cache policy | Generate fixed cache |
| Edit fit protocol | 1.0.0 |
| Grounding resolve | 768 pixels |
| Timestep sampling | krea2_shift |
| Discrete flow shift | 2.5 |
| Text encoder | Qwen3-VL 4B, BF16 |
| VAE | Qwen Image VAE, BF16 |
| Seed | 42 |
Usage Notes
You must use the ComfyUI-Krea2Edit node pack. This will not work with a standard Krea 2 workflow.
Your source image and output image should be the same aspect ratio / resolution.
The base prompt should always be included:
super-resolve the image
Refer to the Krea 2 Identity Edit documentation for background on the image-edit workflow and its conditioning requirements. Its recommended inference settings describe the upstream identity-edit model, not validated settings for this super-resolve LoRA.
As a beta generative super-resolve model, outputs should be checked against the source image: added detail should not be assumed to be a faithful recovery of the original. A tested inference workflow, recommended LoRA strength, and supported upscale factors are not yet documented here.
Credits
- conradlocke/krea2-identity-edit: Krea 2 identity-edit resource used in creating this model.
- lbouaraba/krea2edit-trainer: Training script adapted for use with Musubi Tuner.
- Musubi Tuner: Training framework used for the adapted script.
- zhang0jhon/Aesthetic-4K: Source dataset for the 1,000 fuzzed training images.
- Krea 2 Raw: Base model identified in the training configuration.
Licensing
This model card retains the repository's existing MIT license metadata. The upstream Krea 2 base model and Krea 2 Identity Edit weights are subject to the Krea 2 Community License; this repository's metadata does not replace those terms. Review the upstream licenses before using or redistributing the model. The trainer is published under Apache-2.0, and Aesthetic-4K is listed under MIT.