5.7 KiB
Render effort tiers
Status: newly adopted (October 2026) and on trial. The scheme below was agreed for the next major town batch and is being evaluated there. Everything in it is a proposal with reasons, not a law. Use it, measure it, and adjust the numbers to your own table, budget and renderer.
The one test
Tier = the probability that the place comes into focus at the table.
Not its size, not its importance in the world, not how pretty it would be. A third-rate pub that the storyline sends the party to is tier 1. A master weaponsmith in a town the party only passes through is tier 3. The categories below are examples of where places usually land, never rules.
| Tier | Meaning | Typically | Render deliverables | Budget (one hosted edit model, 2026 prices) |
|---|---|---|---|---|
| 1 | Guaranteed to come into focus | town overviews and approaches; the party's base, inn or hall; the storyline's buildings; the guild's tavern | block model + aerial and approach plates; the full time x weather matrix; every room the party will stand in; scene beats; named staff as reusable anchor characters | ~100-200 renders, ~$10-20 |
| 2 | Likely | adventuring shops (gold, gems, weapons, armor, alchemist, healer); civic and sacred landmarks (temple, moot hall, gates, market, docks) | exterior A/B day + night + one snow; one counter interior; the proprietor as a scene character; goods at the wealth tier; a price list | ~20-40 renders, ~$2-4 |
| 3 | Possible | third-rate pubs, mundane townsfolk shops | one exterior day, one interior, a keeper from the pool | ~5-10 renders, under $1 |
| 4 | Unlikely / background | everything else on the map | a name, a tagline, one line of text, maybe a sign emblem | no renders |
Special-case venues may warrant their own budget regardless of tier; decide that per campaign.
Deliverables per tier, spelled out
Tier 1
- Geometry: block model of the site in its town context; template plates for every camera (approach, aerial, entrance, interior rooms).
- Exterior matrix: Day / Sunset / Night x Dry / Rain / Snow (light) for the A and B cameras (18 frames), plus whatever storm and deep-snow cells your calendar needs (references/matrices.md).
- Interiors: every room the party will stand in, as an empty master (people-free, so it stays reusable as a plate).
- Population x time matrix (venues with people): 4 x 3, Morning / Afternoon / Evening / Late Night x Quiet / Moderate / Packed, from one "reveal from the door" camera (12 frames). A 4 x 4 with an Empty column only for storyline hubs where an empty room is itself a scene (16 frames).
- Scenes: the beats the story needs, on the canon rooms with the canon faces.
- Named staff as reusable characters with locked faces.
Tier 2
- Exterior A and B: Day Dry, Night Dry, one Snow (light).
- One counter interior, empty master.
- The proprietor as a scene character (a face that recurs, not a full anchor pack).
- Goods rendered at the wealth tier; a price list in text.
- Population matrix 2 x 3: Day / Sunset x Quiet / Moderate / Packed (6 frames), for any venue with a counter or benches.
Tier 3
- One exterior, Day Dry. One interior. A keeper drawn from the generic pool.
- Population matrix 1 x 2: its peak hour x Quiet / Packed (2 frames).
Tier 4
- Text only: name, tagline, one line. A sign emblem if it is cheap. No renders.
Model-cost tiering (which agent does what)
Image QA is the real token cost, not the renders. So the agent that executes a tier is sized to the tier:
| Tier | Planning | Execution | Review |
|---|---|---|---|
| 1 | the most capable model writes the brief and reads the report | a mid-sized model runs the lane | a short review by the capable model |
| 2 | brief from a template | mid-sized model | none beyond the lane's own gates |
| 3 | template | a small model | none |
| 4 | — | a local or text-only model (it never has to see an image) | none |
Canon guard for lanes run by smaller models: every brief says "anything not in this brief is marked PROPOSED". A lane invents less when it knows invention is allowed but must be labeled. (The "no glass anywhere" rule that haunted one campaign for a week was a lane's invention; see references/prompting.md on glass.)
Rules
- Play promotes, never demote. A place the players actually enter moves up a tier for the next session. Nothing rendered is ever deleted; superseded files are moved aside.
- Per-tier ceiling. A lane may not exceed its tier's budget without the GM's say-so. Cap retakes and the number of images reviewed, not just renders.
- Attention is the scarcest resource, then tokens, then wattage. Tier budgets cap tokens and renders, but never by pushing review load onto the human: spend more on self-gating so fewer, better frames reach them.
- Canon-direct delivery. Work that passes its own gates goes straight into the campaign's canon folder with a one-line note. A human review entry is reserved for: the face of a story-critical character; a human gate before something permanent (a model fine-tune, a bake); something the GM asked to see. Reviewed means done: review remarks are records, not new work orders.
- Renderer policy is unchanged by tier. Cut spend by volume (fewer takes, re-roll caps, tighter candidate sets), never by switching to a weaker renderer for the cheap tiers. A weaker renderer costs more in retakes and review.
What to evaluate after the trial
- Did the smaller-model lanes invent canon? How often?
- Retake rate per tier compared with capable-model lanes.
- Actual tokens and renders per tier against the budget.
- Did the tiering match what the table actually visited? (That is the only question that matters.)