The build in this episode is a hardware one: an old phone becomes a camera that shows the world through an image model instead of through glass. Every frame you point it at comes back restyled.
The interesting part is not the phone. It is what MOSS Image has to do to make that loop feel live, and where the illusion breaks.
#
The five workflows
| # | Workflow | Role in the loop | | 01 | Colour-depth LoRA | Light and colour control on the restyle | | 02 | 50-step vs 8-step DMD vs 4-step Turbo | Picks the operating point | | 03 | Base vs SFT, same parameters | Picks the model variant | | 04 | JSON Layout structured composition | Keeps the scene structurally readable | | 05 | SFT 50-step high quality | The quality ceiling, used for stills |
#
The step-count decision is the whole design
Live-style rendering has a latency budget measured in single-digit seconds. That rules out 50 steps immediately. Between 8-step DMD and 4-step Turbo, the trade is:
4-step Turbo — fast enough that pointing the phone feels responsive. Detail is soft; fine structure (text, railings, thin branches) dissolves. #
8-step DMD — noticeably richer, but the added latency makes the phone feel like a camera with a delay, not a lens.
The episode picks Turbo for live use and switches to 8-step for “ and capture” moments. That split is the actual engineering answer.
#
Where the illusion breaks
Three things break it, in order of severity:
Motion. Fast panning produces smeared frames because the model re-imagines each frame independently. There is no temporal coherence between frames. 2. Text. Any signage in view gets re-lettered into nonsense within a second or two. 3. Faces. Strangers in the background become stylised, which reads as uncanny rather than magical.
Knowing that list is more useful than any demo reel: it tells you what to point the thing at.
#
What this is actually good for
- Location scouting with a stylised preview — you see composition and lighting relationships quickly.
- Live performance visuals where latency is tolerable.
- Teaching: showing people the difference between “what the model sees” and “what is there” is a better demo than a static image.
#
Files
Bilibili (Chinese original): https://www.bilibili.com/video/BV1rTnN6vE7A/
#
Notes
- The hardware build is described in the video; the repository contains only the ComfyUI workflows.
- Tutorial material, not a product.
- Independent AI-KSK release; not affiliated with the model authors.