This is the first Numonic newsletter. Each week it covers what we shipped, what we tested, and what we learned making AI creative work that remembers how it was made. It’s short, and every reply reaches a person.
Shipped
Numonic Now Answers “How Did I Make This?”
Shipped in the last two weeks:
- Creative sessions for ComfyUI: renders grouped into the session that made them.
- Compare any two generations: prompt, parameters, reference images, and model, side by side.
- Selected, rejected, undecided: on Midjourney chains, mark each generation with a reason. Change your mind and the history keeps both.
- Copy workflow: paste the graph straight back into ComfyUI.
- Composed from: a stitched video shows each part, its frame range, and its own workflow.
Also new: shares that carry the prompt, references, and code; plain-language search (“portraits made with ComfyUI since March 2026”); and large files sent from your AI assistant, for anything over about 3 MB.
Watch
When the Client Asks for One Small Change
A short series on changing one thing in an approved shot and keeping everything else:
- One Small Change: remove an object from a delivered shot, nothing else
- Same Shot, Golden Hour: re-light an approved image without losing it
- Which One Did We Approve? Four near-identical renders, one answer
- Can We Get a Vertical Version? 16:9 to 9:16 without changing the shot (and the 60-second version)
Getting started? Set up the Numonic ComfyUI node · Organize and recreate your ComfyUI workflows · Bring thousands of ComfyUI images in as they are · Connect Claude or any MCP agent · Share your creations
From the Lab
Our Agent Made a Film, and Three People Found What It Missed
Claude Code made “The Paper Crane” for us. Its own checks passed it. Then three of us watched it and saw the crane change within a minute. For the remake, it wrote its rules before generating anything, and rejected 84 of 96 storyboard candidates before a single clip was made.
Read what changed · Watch the film · Walk through every piece, in order
We also stretched an LTX-2.5 video from 30 to 62 seconds. One join failed 17 times out of 18, until we moved it to a closer, busier moment. Read the post · Watch the experiment
Made by agents, on Numonic:
- Latent Fauna, a field guide Claude bound using only what Numonic recorded.
- The Wrong 20th Century, an archive where every artifact cites its parents.
- Orbit Title Card, made over MCP (the emerging standard for AI tool integration) and shared with its code.
Worth Reading
Three Guides and Three Outside Reads
- ComfyUI PNG text chunks: where the workflow, parameters, and seed are stored
- EU AI Act Article 50: AI disclosure rules for creators, updated for the July 2026 Guidelines. If you deliver AI images to clients in the EU, start here.
- ComfyUI LoRA manager, folders, and metadata: a practical guide
And from elsewhere:
- StreamEval tested 37 video models on 300 scripted cases. Good-looking, identity-preserving output does not always mean the requested change was made. That is the split we use in the One Small Change series: did it make the change, and did it keep everything else?
- Code2Video-AA had 27 models write short product videos as HyperFrames HTML. Code should make a changed word or number a small edit. The benchmark does not measure that, so test it on your own brief.
- The MOSS-Image-Edit-HQ model card recommends one checkpoint, logs where the newer one got worse, and says the 8-step mode is for drafts. Pin the checkpoint and the workflow together.
Your Turn
What Would You Want to See?
If an agent made a film for you in Blender, ComfyUI, or Unreal, what would you want to see about how it was made? Reply to this email. We read every answer, and the best ones shape what we build next.