The #535 QR lane launched on monster-labs/control_v1p_sdxl_qrcode_monster (SDXL, v1 — the only SDXL release). First real-world results are visibly below monster-labs' showcase gallery, which is generated by their SD 1.5 Monster v2 — the widely-acknowledged best art/code blender; the SDXL port trades that blending for base-render sharpness. No newer release (no SDXL v2, no FLUX) exists.
Plan:
artpipe: add a variant param to qr_monster — classic (SD 1.5 base + control_v1p_sd15_qrcode_monsterv2 subfolder, 512/768px, new default) and sdxl (existing stack, 768/1024px). One variant resident at a time with VRAM eviction on switch (9500 MB lease covers either stack, not both). Per-variant tuned defaults: classic scale 1.1 / 36 steps, sdxl scale 1.35 / 32 steps; guidance 7 both. Size requests snap to the variant's native range.
backend/angular: plumb variant through GenerateQrArtCommand + the QR tool UI; temporarily raise the QR endpoint rate limit to 60/hr (1/min) for the tuning/testing window (was 6/hr).
infra/CI: env-image rebuild (SD1.5 base fp16 + v2 ControlNet ≈ +3.5 GB), redeploy, then an on-hardware tuning pass (scannability vs aesthetics) before locking defaults.
Watch: root repo of the SD15 ControlNet is the WEAKER v1 — weight patterns must pin v2/* (regression-guarded in tests).
The #535 QR lane launched on `monster-labs/control_v1p_sdxl_qrcode_monster` (SDXL, v1 — the only SDXL release). First real-world results are visibly below monster-labs' showcase gallery, which is generated by their **SD 1.5 Monster v2** — the widely-acknowledged best art/code blender; the SDXL port trades that blending for base-render sharpness. No newer release (no SDXL v2, no FLUX) exists.
**Plan:**
- artpipe: add a `variant` param to `qr_monster` — `classic` (SD 1.5 base + `control_v1p_sd15_qrcode_monster` **v2 subfolder**, 512/768px, new default) and `sdxl` (existing stack, 768/1024px). One variant resident at a time with VRAM eviction on switch (9500 MB lease covers either stack, not both). Per-variant tuned defaults: classic scale 1.1 / 36 steps, sdxl scale 1.35 / 32 steps; guidance 7 both. Size requests snap to the variant's native range.
- backend/angular: plumb `variant` through GenerateQrArtCommand + the QR tool UI; **temporarily raise the QR endpoint rate limit to 60/hr (1/min) for the tuning/testing window** (was 6/hr).
- infra/CI: env-image rebuild (SD1.5 base fp16 + v2 ControlNet ≈ +3.5 GB), redeploy, then an on-hardware tuning pass (scannability vs aesthetics) before locking defaults.
- Watch: root repo of the SD15 ControlNet is the WEAKER v1 — weight patterns must pin `v2/*` (regression-guarded in tests).
Shipped, deployed, and verified end-to-end 2026-07-28 ~21:00Z.
artpipe #45 + #46 merged: classic (SD1.5 Monster v2) variant as default, scan-pyramid verification, tuned defaults. On-hardware tuning results: classic ≤1.3 never decodes; ≥1.5 decodes 100% of the seed-fixed grid — but only at 512px downscale, never at the raw 768px render (the artistic texture reads as noise to zxing until the module pyramid collapses, i.e. exactly how phones scan these). Hence: default scale 1.5 + a downscale/grayscale decode pyramid in the verifier (also benefits sdxl).
backend #504 + angular #599 merged & deployed: variant plumb, Classic|SDXL toggle live on the tools page, rate limit 60/hr for testing (restore tracked in #890).
Final verification on the deployed worker image (digest e1b75e74), pure defaults: scannable=true, attempts=1, scale_used=1.5, gallery-quality output.
Closing. — Opus 5 Agent
Shipped, deployed, and verified end-to-end 2026-07-28 ~21:00Z.
- artpipe #45 + #46 merged: classic (SD1.5 Monster v2) variant as default, scan-pyramid verification, tuned defaults. **On-hardware tuning results:** classic ≤1.3 never decodes; ≥1.5 decodes 100% of the seed-fixed grid — but only at 512px downscale, never at the raw 768px render (the artistic texture reads as noise to zxing until the module pyramid collapses, i.e. exactly how phones scan these). Hence: default scale 1.5 + a downscale/grayscale decode pyramid in the verifier (also benefits sdxl).
- backend #504 + angular #599 merged & deployed: variant plumb, Classic|SDXL toggle live on the tools page, rate limit 60/hr for testing (restore tracked in #890).
- Final verification on the deployed worker image (digest e1b75e74), pure defaults: **scannable=true, attempts=1, scale_used=1.5**, gallery-quality output.
Closing. — Opus 5 Agent
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
The #535 QR lane launched on
monster-labs/control_v1p_sdxl_qrcode_monster(SDXL, v1 — the only SDXL release). First real-world results are visibly below monster-labs' showcase gallery, which is generated by their SD 1.5 Monster v2 — the widely-acknowledged best art/code blender; the SDXL port trades that blending for base-render sharpness. No newer release (no SDXL v2, no FLUX) exists.Plan:
variantparam toqr_monster—classic(SD 1.5 base +control_v1p_sd15_qrcode_monsterv2 subfolder, 512/768px, new default) andsdxl(existing stack, 768/1024px). One variant resident at a time with VRAM eviction on switch (9500 MB lease covers either stack, not both). Per-variant tuned defaults: classic scale 1.1 / 36 steps, sdxl scale 1.35 / 32 steps; guidance 7 both. Size requests snap to the variant's native range.variantthrough GenerateQrArtCommand + the QR tool UI; temporarily raise the QR endpoint rate limit to 60/hr (1/min) for the tuning/testing window (was 6/hr).v2/*(regression-guarded in tests).Implementation up:
Remaining: env+worker image rebuild → redeploy → on-hardware tuning pass before locking defaults. — Opus 5 Agent
Shipped, deployed, and verified end-to-end 2026-07-28 ~21:00Z.
Closing. — Opus 5 Agent