docs: RD sprite-ideation FINDINGS + memory — exploration wrapped, canon ratified, art direction set

Base-only M03 recipe, 64/48/32 sprite budget, style-LoRA reservation
for tiles/icons. Art direction = watercolor scenery (waterland-dev) +
pixel sprites; ratified to advisor CONCEPT.md 3eeeaf4.
This commit is contained in:
2026-08-14 22:27:36 -07:00
parent 3ed3a2d0ee
commit 21cca5cbdb
2 changed files with 85 additions and 11 deletions
+18 -11
View File
@@ -63,17 +63,24 @@ _As of 2026-08-07 (post-falsifier):_
model's floor** (validated A/B on booth rd-validation).
- `[2026-08-15]` **Style LoRAs recovered — operator decrypted all 24
himself (with permission); now live in ComfyUI.** Genuine SD1.5
LoRAs (kohya, 792 tensors each), symlinked `models/loras/rd/`,
enumerated by LoraLoader. Validation step 2 passes (rd/topdown @1.0,
delta 41.7 vs baseline; booth rd-loras). **Licensing (operator
directive): weights are copyrighted, licensed-for-use,
NOT-for-export — fleet-internal only, no redistribution/publish/
serve.** Ideation output stays `_ph` per asset policy regardless.
**Trigger words are per-LoRA and non-guessable** (gameboy64, gba64,
pixelflat64, …) — 20 of 24 carry one; full map catalogued on
comfy-dev, requested for local record. First ideation batch (M03
sprite-grid 32/48px @512-downscaled, grunt/burster) gated on
operator go.
LoRAs (kohya, 792 tensors each). **Licensing (operator directive):
copyrighted, licensed-for-use, NOT-for-export — fleet-internal
only.** Full trigger map at `tools/retro-diffusion/lora-triggers.md`.
- `[2026-08-15]` **M03 sprite-ideation exploration COMPLETE (2 batches)
— pipeline proven, findings ratified to canon, exploration wrapped
(operator).** Key results (full record: `tools/retro-diffusion/
FINDINGS.md`): base `rd_pixel_v14` alone renders M03 + mismatched
eyes; style LoRAs kill the hero (object priors) → reserved for
tiles/icons/env props; **sprite budget 64px full-body / 48px bust /
32px blob**. Booths m03-batch01 / m03-batch02 (ephemeral).
- `[2026-08-14]` **Art direction resolved (operator): watercolor
backgrounds/scenery + pixel-art sprites.** Watercolor scenery
pipeline owned by **waterland-dev** (fleet peer). Pixel sprites =
local RD base ideation. Ratified to advisor CONCEPT.md (§ Tech & art
direction + decisions log + open-Q #2 sprite-grid struck) at
`3eeeaf4`. Ideation-grade — shipped art still human-authored, M03 ref
stays locked syn02/syn04b. Loop-in waterland-dev when scenery work
actually starts (no ask ripe yet).
## Recent decisions
+67
View File
@@ -0,0 +1,67 @@
# Retro Diffusion — M03 sprite-ideation findings
Durable record of what the batch-01/02 ideation runs established about
driving the local RD stack for U R M03 sprites. Design canon for the
art direction lives in the advisor repo (CONCEPT.md § Technology & art
direction); this is the implementation-side how-to.
**Status: IDEATION / pipeline proof.** Shipped sprites are human-authored
(asset policy). M03's canonical visual reference is locked (syn02/syn04b,
advisor repo). These findings say how to make RD *propose* sprites, not
what ships.
## Art direction (operator, 2026-08-14)
**Watercolor backgrounds + scenery, pixel-art sprites.** Scenery pipeline
is owned by **waterland-dev** (watercolor generation stack). Sprite
ideation is the local RD base checkpoint, below. Per-asset direction dials
in later.
## The M03 sprite recipe
- **Base `rd_pixel_v14` + repaired `rd_vae_v14`, NO style LoRA.** On this
checkpoint the un-LoRA'd base is the only config that holds M03 as a
character; every style LoRA imposes an object/subject prior that
displaces the subject (character LoRAs → fantasy battle units / cyclops;
flatshading → lanterns; frontfacing+trigger → pots; 1bit → abstraction).
- **Render 512×512, downscale nearest-neighbor** to the target budget
(512 is above the model's floor; 128 direct is mush).
- **Foreground wanted features positively; do not rely on negatives** — a
strong prior beats a negative on this checkpoint (cfg 7.0). The base's
own weak negative let `pedestal`/`busy background` through repeatedly.
- Signature detail first in the prompt (objects, never abstractions/animal
similes — "column legs on flat pads", not "elephant feet").
## Sprite budget (resolves the sprite-grid open question)
Readable-detail floor tracks feature-size-relative-to-frame, not a fixed
number:
- **64px** — full-body sprite; silhouette + mismatched-eye read both hold.
- **48px** — bust/head-dominant crop only; a full-body sprite loses small
features (recessed lens dissolves).
- **32px** — blob-squint silhouette only.
## Mismatched eyes
Diffusion CAN render them — the base drew a large pale lens beside a small
dark recessed lens unaided at 64px (batch 02, `I_s1004`/`I_s1001`). So the
eyes are NOT forced to be hand-authored by a pipeline limitation (though
MOE.md's "authored, not procedural" may still hold them by choice).
## Style-LoRA reservation
The 24 style LoRAs are counterproductive for the hero character but are
the right tool for **tiles / icons / environment props**`tiling*`,
`gameicons`, `uipanel`, console-palette LoRAs — where an object/style
prior is a feature. Reserved for whenever environment-prop work starts
(scenery itself is watercolor via waterland-dev, not RD).
## Trigger words
See `lora-triggers.md` (this dir) for the full per-LoRA trigger map.
## Batch briefs
`batch-01-brief.md`, `batch-02-brief.md` (this dir) — the run specs and
their booth results (booths ephemeral, ~24h).