From b9b19f7ecb9b8549986e2f6bd1fcec3914b68882 Mon Sep 17 00:00:00 2001 From: "Danilo M." Date: Fri, 30 Oct 2026 18:57:57 +0100 Subject: Record upscaler test results and imajev prompt rule Three test images (clean portrait, landscape, noisy phone JPEG) at 4x picked the foolhardy 4x model for clean photos, RealESRGAN_x4plus for degraded ones and the anime model for art. The suggestion prompt must weigh source quality, since imajev chose a detail model for the noisy photo. Co-Authored-By: Claude Opus 5.5 --- AGENTS.md | 12 ++++++++++++ docs/superpowers/specs/2026-10-30-upscale-qt-design.md | 5 +++++ 2 files changed, 17 insertions(+) diff --git a/AGENTS.md b/AGENTS.md index d196e48..4f98245 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -22,6 +22,18 @@ Design spec: `docs/superpowers/specs/2026-10-30-upscale-qt-design.md`. --image -p `. Its answer is free text and gets matched to a filename from the upscalers dir; never trust it as a path. - Both run through `QProcess` so the UI never blocks. +- `sd-cli` loads ESRGAN (RRDBNet) and MoSR models only, detected from the + weights. SwinIR, HAT, DAT, SPAN, Compact and face restorers do not load. +- Model choice, measured on 4x crops of a clean portrait, a landscape and a + noisy 2009 phone JPEG: the 4x foolhardy model is best on clean photos + (people and nature), `RealESRGAN_x4plus` on noisy or compressed sources + (smooth, slightly waxy skin), `RealESRGAN_x4plus_anime_6B` for anime and + art only. Rejected after testing: 4xNomos2_hq_mosr (amplifies JPEG + blocks), 4xNomos8kSC (artifacts on noisy input), 4xLSDIRplusR (plastic + on clean input), 4xLSDIRplusC (close to the foolhardy model, not better). +- Timing: an RRDBNet model takes about 25 s per megapixel of input, so a + 10 MP photo takes about 4 minutes and outputs a 160 MP image. imajev + answers in about 7 s. ## Conventions diff --git a/docs/superpowers/specs/2026-10-30-upscale-qt-design.md b/docs/superpowers/specs/2026-10-30-upscale-qt-design.md index 8db80a5..ffcb5da 100644 --- a/docs/superpowers/specs/2026-10-30-upscale-qt-design.md +++ b/docs/superpowers/specs/2026-10-30-upscale-qt-design.md @@ -42,6 +42,11 @@ job. Batch mode later adds UI over the same list. 2. If imajev is configured, run `llama-mtmd-cli -m --mmproj --image -p `, the prompt listing the upscaler filenames and asking for exactly one back. + The prompt also tells it to weigh source quality (noise, compression, + blur) and not only the subject: in testing it picked a detail model for + a noisy phone photo where a denoising one did far better. If that keeps + happening, add an optional `models.txt` (filename plus one-line note) + fed into the prompt. The answer is matched to a filename (exact filename, then stem, then case-insensitive substring); a match preselects the dropdown and the status bar says "imajev suggests X". No match or failure: dropdown -- cgit v1.2.3