A graph you can rerun: load the model, apply the trained files, decode, write the stills.
Client
Under NDA
Disciplines
Generative AI, ML ops
Stack
ComfyUI, Flux, safetensors
Engagement
Pipeline under NDA
The client needed images that stayed on-model: the same subject, the same finish, in volume. A public generator with a new prompt each time would not hold a brand or a character.
They also needed the ability to train - to add tensor files that pull the base model toward their own material - and to run that training and generation as a graph, not as a one-off notebook.
Open models are easy to demo and easy to drift. The work is making a graph that other people can run next week.
We built the path as a ComfyUI workflow: loaders for the diffusion model, VAE and trained tensors, then sampling and decode. Weights live as safetensors so the same files can be versioned and moved.
Training is a separate step that produces files the graph consumes. Generation does not retrain on every still.
A repeatable generation graph and the trained tensor files the client actually runs. Operators change the prompt and the batch, not the plumbing.
The same constraint - one approved look, many outputs - is the production version of what the AI image generation service describes.
The engagement is under NDA. We do not publish the client's visual IP, training set, or output volumes.
The graph is the client's generation path. New tensor files can be swapped in when the look needs to move.
This engagement is covered by a confidentiality agreement, so the client is not named and some implementation detail is described in general terms.
Related solutions
