90bd06c328
Round-2 close-out: all lane work landed on main (C keeper-stand re-land, B streetscape, F integration), stray lane-c/keeper-stand branch deleted (content identical), QA strict green. ROUND3_INSTRUCTIONS.md assigns the polish/validation round: F leads (manifest + hours + QA gates), B shaves the spawn draw budget, C+E validate the GLB path via a local depot mode, A verifies one stale finding, D supports. priorwork/ archives the original per-lane prompts. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
23 lines
2.5 KiB
Plaintext
23 lines
2.5 KiB
Plaintext
e cut off
|
|
Good call — let me check what MODELBEAST is and whether MeshGod's backend is even fal/Hunyuan, before anyone pays for cloud gen.
|
|
|
|
MODELBEAST is the answer to your instinct — it's your **local-first M3 Ultra rig** (`100.89.131.57:8777`) with on-device image→GLB operators (`sf3d`, `trellis_mac` on MPS) that cost only electricity. Let me check what's actually live and how it differs from the cloud path.
|
|
|
|
MODELBEAST is **live** and has exactly what we need — **local** operators: `flux_local` (text→image on M3 Ultra, for facades), plus `sf3d` / `trellis_mac` (image→GLB on MPS) + `bg_remove_local` (for props). Both cloud costs are avoidable. Let me check which local operators are actually *ready* (weights installed) vs gated.
|
|
|
|
Now I have the real picture. Key facts:
|
|
- **`flux_local`** uses `flux2-klein-4b` = **Apache/ungated, works out of the box** → facades can be **100% local, free**.
|
|
- **`sf3d`/`trellis_mac`** (image→GLB, local, free) are **HF-gated** — need weights set up.
|
|
- The fal cloud alternative is `fal_hunyuan3d` at **~$0.16/model (+$0.48 for textures)** — so the real cloud price is higher than the "30¢" estimate.
|
|
|
|
Let me check whether I can drive MODELBEAST headless (token/host) and confirm the local generators are actually ready.
|
|
|
|
**I'm physically on the m3ultra box** (IP 100.89.131.57, `m3ultra.local`) — MODELBEAST is running right here on localhost, and I'm the owner. So I can drive the local operators directly (they're just `<venv-python> run.py --input … --outdir … --params …` subprocesses). Let me check whether `flux_local`'s venv + weights are actually installed so I can test one facade locally.
|
|
|
|
**Local facade gen is fully ready** — the `flux_local` venv exists and **FLUX.2-klein-4B + 9B weights are already downloaded** (no 15GB wait, no cloud key). Let me check the exact CLI and generate one test facade to prove the quality matches the anchor before committing to all 23.
|
|
|
|
Local facade gen works — **7.4 seconds, free, on-device**. Let me eyeball the quality against the style anchor.
|
|
|
|
**This is the answer.** Locally-generated in 7.4s, free — and it's *excellent*: photoreal 90s-Australian toy shop, **genuinely blank signboard** (no baked text), primary-colour render, toys in the window. Style-matches the anchor set, arguably cleaner than some djsim ones. So facades are **100% local, free, ~7s each** — zero OpenRouter cost needed.
|
|
|
|
Let me wire this local path into `gen_skins.py` as a first-class `--local` backend, then generate all 23. |