urm03/tools/retro-diffusion/batch-01-brief.md

RD ideation batch 01 — M03 silhouette + sprite-grid probe

Purpose: first Retro Diffusion ideation run for U R M03. Art-direction

• experiment design; comfy-dev executes on ComfyUI (irv-ml1). Output is _ph ideation ONLY — informs the eventual art pass, does NOT enter the greybox (greybox stays art-free per canon).

Canon source: M03 design law, MOE.md § Design language (advisor repo). Cues below are the drawable-object translation of that law.

What this batch answers

1. Does RD's pixel-art base render our gentle giant at all — the blob-squint test at pixel scale (round + huggable + the signature asymmetries), distinct from the menacing enemy read.
2. Sprite-grid-size hypothesis (parked playtest question, CONCEPT.md §Open-Q #2 / ROADMAP): at what pixel budget (32 / 48 / 64) does M03's silhouette AND the mismatched-eye read survive? This is the real deliverable — a downscale ladder we can eyeball.
3. House-style pick: which character LoRA carries M03 best — gamecharacters (duelystchar), gamecharactersretro (no trigger), or gamecharactersanime (no trigger).

Model config

• Base: rd_pixel_v14 + rd_vae_v14 (the repaired VAE).
• Render at 512×512, downscale nearest-neighbor to 32/48/64 (model resolution floor — 128 direct is mush, validated).
• LoRA strength start 1.0; sweep 0.6–1.0 if style overpowers structure.
• SD1.5 family: negatives ARE available here (unlike the flux/Arbo path) — use them. Trigger word leads the prompt.
• comfy-dev owns sampler/steps/CFG and will content-check every render (size + brightness/variance, eyes on pixels — not "graph executed").

M03 — the canonical prompt (load-bearing; this IS the canon translation)

Signature detail FIRST (Arbo lesson that transfers), objects not abstractions (no animal similes — "elephant feet" once literally made him an elephant; use "column"/"pad"), the two-object eye framing MOE.md itself validated (plain "mismatched eyes" collapses to symmetry):

Positive:

duelystchar, a gentle giant farming robot, rounded dome head, one dim small recessed lens eye beside one bright glowing bolted-on lamp eye (the bright lamp is the warm one), barrel-shaped body, rounded shoulders, thick cylindrical column legs ending in flat round pads, one arm ends in a watering-spray nozzle, a simple weapon bolted on over the body, warm, huggable, plain background, full body, pixel art

Negative:

humanoid feet, boots, sleek, angular, military, menacing, cyclops, single eye, symmetrical matching eyes, human face, text, watermark, busy background

Judge against MOE.md's own tests: blob-squint (huggable at silhouette), mismatched-eye (patched/bright eye reads different and warmer), bolted-weapons (armed, still passes blob-squint).

Enemies (secondary — tests the menace contrast)

Enemies read menacing by canon — the inverse of Moe's authored softness.

• Grunt (untagged): duelystchar, a small simple rogue robot, blocky, plain olive-drab, no glow, faintly menacing, plain background, full body, pixel art. Neg: cute, friendly, rounded, warm, glowing eyes.
• Burster (HEAT): a round rogue robot with a glowing molten-orange core in its belly, cracked armor, overheating, menacing, plain background, full body, pixel art. Neg: cute, friendly, gentle, human, cool colors.

Batch shape (keep it a probe, not a production run)

• A — M03 silhouette: the canonical prompt, ~6 seeds at 512, LoRA duelystchar @1.0.
• B — house-style compare: one strong seed from A, re-run through gamecharactersretro and gamecharactersanime (no trigger — just load).
• C — grid ladder: best 1–2 M03 renders, downscaled to 32/48/64 shown side by side. This is the sprite-grid-size answer.
• D — enemies: grunt ×2, burster ×2 at 512.

NOT this batch (deliberately deferred)

• ControlNet from our hand-authored maps. Our greybox m03_ph map is deliberately crude; constraining to it would defeat "let RD propose a better silhouette." ControlNet (Sketch/Pose) is the RIGHT tool once we've PICKED a silhouette — for turnarounds/animation frames off an agreed design. Second-pass, not first.
• Tag-hue / VFX exploration, rig frames, Khoe — later batches.

Output

Booth with the A/B/C/D grids; reply on the RD thread to m03-dev with the URL + which LoRA carried M03 best (your read) + anything that fought the prompt. m03-dev takes the operator's verdict from there.

Licensing: weights fleet-internal, not-for-export (operator directive). Output is _ph ideation, operator-judged.