ComfyUI, LoRA experts needed

AI IMAGE PIPELINE - CONTRACT

ComfyUI, LoRA training, ControlNet. Ongoing work.

We're a UK educational publisher producing a large illustrated fiction list for schools. We need someone to build and run an image generation pipeline for it.

THE WORK

Each book series has its own visual style, defined by a written style guide and a small set of reference images. Your job is to turn that into a repeatable engine: train the style, lock it, and produce around 110 print-ready images per series to a supplied brief.

The images are composed to a page layout, so composition is directed, not left to the model. Each one has a specified size and a specified job on the page.

WHAT YOU NEED TO BE ABLE TO DO

- Build custom ComfyUI workflows at graph level. Not just run other people's.
- Train LoRAs properly. Dataset curation, captioning, hyperparameters, and knowing when a model has overfit.
- Control composition with ControlNet, depth, pose and regional prompting.
- Inpaint and repair to a print standard.
- Script and batch with Python. This is volume work and it cannot be done by hand in a GUI.
- Prepare for print: 300dpi, upscaling, CMYK, colour management.
- Document what you build, so someone else can run it.

WHAT WE'RE NOT LOOKING FOR

- Midjourney-only. We need training and control, not a closed system.
- Prompt engineering as a job title.
- A beautiful portfolio of one-off images. One stunning picture tells us nothing about producing 110 consistent ones.

THE TEST

Paid. We send you a style guide and five reference images. You send back six images in that style: a rope, a hand holding a coin, a boat, and three we'll name.

We are not judging how good image one is. We are judging whether image six matches it.

STRUCTURE

First engagement is fixed scope: build the workflow, prove it on one series, document it. Hourly/daily work follows once the machine exists and we know what it needs.

RATES

- Free simple test:
- Build: fixed fee,
- Hourly or daily rate thereafter:

TO APPLY

Send:

1. A workflow you have built, and a short note on what problem it solved
2. A LoRA you have trained, with the dataset approach you used
3. A set of images showing one style or character held across at least six outputs