Project Garden Lens
Hello everyone,
I'm a professional horticultural photographer and have spent decades photographing gardens, plants, and flowers. Over the years I've accumulated an extensive archive of original RAW files paired with professionally retouched final TIFFs.
Many of the edits involve:
• Leaf blemish removal
• Insect damage repair
• Petal cleanup
• Foliage refinement
• Distraction removal
• Selective focus enhancement and visual attention control
What has me curious is whether an archive like this could be used to train or fine-tune a computer vision model to perform a first-pass botanical retouching workflow.
My goal wouldn't be to replace professional judgment. I'd simply love to reduce some of the repetitive cleanup work so I can spend more time behind the camera and less time in front of Photoshop.
Has anyone worked with a similar before-and-after image dataset for image enhancement or retouching applications?
Does this seem technically feasible, and if so, what type of model or workflow would you investigate first?
I'd appreciate any thoughts or suggestions.
Thank you!
Doreen
Eye of the Lady
You do have useful data for sure! And the most valuable would be if you have before/after pairs of images so you can train a model how to fix a blemish as you would expect.
If you are truly just trying to take care of “basic” repairs, the easiest path is to play with the foundation models in Photoshop, ChatGPT, Gemini, etc. You could circle the spot to repair and ask it to try.
The more robust path would be combining a few different tools to accomplish your goal at a more precise level. I’ll put the ChatGPT response below for completeness, but here is how I would summarize it:
- Data: before/after pairs of images, as well as annotations showing where the repairs are needed
- Roboflow: train on “before” images to find areas to repair
- Hugging Face Diffusers: fine tune the Stable Diffusion XL model (using LoRA) with your original image, original with a white mask on the blemish, and after image showing a correct repair
- Actual production runs - Roboflow finds the problem area and puts a white mask on it, then your trained SDXL model repairs the white mask area given the training you taught it
Major project but doable! Best of luck with it all!
Per ChatGPT:
Roboflow could be useful, but probably as the masking/detection part rather than the pixel-repair part. You could use Roboflow to label and train a segmentation model that finds leaf blemishes, insect damage, damaged petals, or distracting background elements. Then you’d pass those masks to an inpainting/restoration model to generate the repaired pixels.
For the before/after learning, I’d look more at diffusion/image-to-image tools: Hugging Face Diffusers, Stable Diffusion inpainting, InstructPix2Pix-style fine-tuning, or a ComfyUI prototype. ComfyUI would be good for experimenting visually; Diffusers/PyTorch would be better for training and batch processing.
I’d start by testing Photoshop/Lightroom generative remove tools and a ComfyUI inpainting workflow on 50–100 representative images. If generic foundation models get 70–80% of the repetitive cleanup right, maybe no custom training is needed. If they miss botanical realism or your personal retouching taste, then the paired archive becomes valuable for fine-tuning.