18h • 📚 Resources & Finds
I've been testing image models since the local SD era. AMA.
Most people met image gen after chat made it mainstream.
I met it earlier, on local hardware, with Stable Diffusion, when the job was still: make it run, make it hold, make it fail honestly.
That history is not a flex. It is a testing habit.
The yardstick
When I try a new image model, I do not ask "is it pretty?"
I ask:
- Does a tight inventory hold (counts, props, who is in frame)?
- Do hands, materials, and grit survive, or does it go plastic?
- Is the light documentary, or beauty-filter default?
- Can I lock a still before I trust motion?
Pretty without a yardstick is noise.
The method (still first)
What I run, even on cloud models now:
- Subject + action first. Not camera brand names.
- Fill the frame with the work (centre subject, intentional crop).
- Mechanics checklist (hands on tools, cables, sandbags, real set mess).
- Light as production light, not influencer key.
- Materials named (dust, tape, fabric, metal) so the model cannot invent void.
- One craft lock (film still, grain, matte blacks) plus a hard forbid list.
- Still first. Motion second. Never use video to fix a weak plate.
If the still fails dual judgment, I re-roll or refine on one axis. Wording or seed or a small edit. Not all three at once.
Proof attached
For this post I ran that same discipline on a brief I care about:
Behind the scenes of a film being recorded. Ultra-wide 21:9, centre-framed subject. Still locked, then motion from that plate only.
Not decoration. A live test of whether the stack holds a real set.
Floor is yours
Ask me anything on:
- Prompting for messy real environments (sets, hands, practicals)
- Workflows still to motion, local era to cloud era
- Gen judgment how you know something is good enough to ship
I'll batch replies this week.
No wrong questions.
Specific beats welcome ("how do you kill plastic skin?", "what do you ban by default?").
More of my work: https://pushingsquares.com/ari
//A<3
9
10 comments
I've been testing image models since the local SD era. AMA.
skool.com/cliefnotes
What we give away free beats most paid courses. Build durable AI systems with a Marine vet and Edinburgh researcher. 40+ lessons, growing.
Leaderboard (30-day)
1
+1392
2
+1303
3
+1227
4
+641
5
+594
Powered by