LZO-1-Preview(experimental)

LZO-1-Preview (Lossless-Zoom-Operator) is an experimental adapter for black-forest-lab’s FLUX.1-Kontext-dev. It is an experimental LoRA designed to zoom into a defined object frame within an image without altering the object's position, maintaining strict center-staged positioning. The model was trained on 550 image pairs (275 original “start” images and 275 “end” images). Synthetic result nodes were generated using Gemini 2.5 Flash Image Preview from Google and annotated with DeepCaption-VLA-7B.

[photo content], zoom in on the specified [face/object/region close-up], enhancing resolution and detail while preserving sharpness, realism, and original context. Maintain natural proportions and background continuity around the zoomed area.

Sample Inference

[photo content], zoom in on the specified [face close-up], enhancing resolution and detail while preserving sharpness, realism, and original context. Maintain natural proportions and background continuity around the zoomed area.

Example 1	Example 2

This is an experimental model and may generate sub-optimal results at times. Make sure the uploaded image is of good quality, as poor-quality input images may produce artifacts. Also, ensure that the input includes a proper prompt for optimal results. The model is optimized and works well on human face cards, anime images, automotive images, and sports action images.

---

Parameter Settings

Setting	Value
Module Type	Adapter
Base Model	FLUX.1 Kontext Dev - fp8
Trigger Words	[photo content], zoom in on the specified [face/object/region close-up], enhancing resolution and detail while preserving sharpness, realism, and original context. Maintain natural proportions and background continuity around the zoomed area.
Image Processing Repeats	50
Epochs	30
Save Every N Epochs	1

Labeling: DeepCaption-VLA-7B(natural language & English) 

Total Images Used for Training : 550 Image Pairs (275 Start, 275 End)

Synthetic Result Node generated by gemini-2.5-flash-image-preview

Training Parameters

Setting	Value
Seed	-
Clip Skip	-
Text Encoder LR	0.00001
UNet LR	0.00005
LR Scheduler	constant
Optimizer	AdamW8bit
Network Dimension	64
Network Alpha	32
Gradient Accumulation Steps	-

Label Parameters

Setting	Value
Shuffle Caption	-
Keep N Tokens	-

Advanced Parameters

Setting	Value
Noise Offset	0.03
Multires Noise Discount	0.1
Multires Noise Iterations	10
Conv Dimension	-
Conv Alpha	-
Batch Size	-
Steps	3900 (Low(700))
Sampler	euler

---

Trigger words

You should use [photo content] to trigger the image generation.

You should use zoom in on the specified [face/object/region close-up] to trigger the image generation.

You should use enhancing resolution and detail while preserving sharpness to trigger the image generation.

You should use realism to trigger the image generation.

You should use and original context. Maintain natural proportions and background continuity around the zoomed area. to trigger the image generation.

Download model

Download them in the Files & versions tab.