feat(v2v): real multi-strength Right variants for the forest scale (vertical slice) #11

Merged
benstull merged 3 commits from feature/v2v-forest-right-variants into main 2026-06-08 13:58:54 +00:00
Owner

First real-content slice of the offline v2v pipeline (operator chose a vertical slice, session 0012) — so the dreamlike Right axis can finally be judged by eye on real footage, breaking the 'can't judge the ring liked without real content' chicken-and-egg.

What shipped:

  • simulator/bake_right_variants.py — the session-0008 POC flow-restyle algorithm (SD img2img keyframes + Farneback optical-flow tweens = temporally calm, no boiling) productionized as a repo-resident offline baker, with a pure, unit-tested strength curve.
  • Ran it for the forest scale: real Right variants 1–4 (replacing 1 real + 3 ffmpeg placeholders).
  • setup_sample_media.py now only stages the neutral base; the baker owns the variants (re-running setup never clobbers a real bake).

By-eye finding (the point of the slice): sd-turbo's painterly transform is sharply non-linear in strength — barely registers <0.5, ramps hard through ~0.5–0.65. A linear 0.28–0.58 ramp left levels 1–3 ~identical to raw, so I re-baked with a clustered 0.46/0.52/0.58/0.64 curve: a real perceptual ramp (gentle haze → clearly impressionist → strongest), verified frame-by-frame. Curve is one-line tunable + re-bake; the operator will likely fine-tune before locking the look.

Deferred (next): real Right variants for cosmos/abyss (need their real strict-PD bases sourced first) and the real i2v ring transitions (need a 2nd real base + a generative-video model not in the POC).

Plan: docs/superpowers/plans/2026-06-08-v2v-vertical-slice-forest-right-variants.md. Tests: +4 (curve). Suite 228 passed / 2 skipped.

🤖 Generated with Claude Code

First **real-content** slice of the offline v2v pipeline (operator chose a vertical slice, session 0012) — so the dreamlike Right axis can finally be judged by eye on real footage, breaking the 'can't judge the ring liked without real content' chicken-and-egg. **What shipped:** - `simulator/bake_right_variants.py` — the session-0008 POC flow-restyle algorithm (SD img2img keyframes + Farneback optical-flow tweens = temporally calm, no boiling) productionized as a repo-resident offline baker, with a pure, unit-tested **strength curve**. - Ran it for the **forest** scale: real Right variants 1–4 (replacing 1 real + 3 ffmpeg placeholders). - `setup_sample_media.py` now only stages the neutral base; the baker owns the variants (re-running setup never clobbers a real bake). **By-eye finding (the point of the slice):** sd-turbo's painterly transform is sharply non-linear in strength — barely registers <0.5, ramps hard through ~0.5–0.65. A linear 0.28–0.58 ramp left levels 1–3 ~identical to raw, so I re-baked with a clustered **0.46/0.52/0.58/0.64** curve: a real perceptual ramp (gentle haze → clearly impressionist → strongest), verified frame-by-frame. Curve is one-line tunable + re-bake; the operator will likely fine-tune before locking the look. **Deferred (next):** real Right variants for cosmos/abyss (need their real strict-PD bases sourced first) and the **real i2v ring transitions** (need a 2nd real base + a generative-video model not in the POC). Plan: `docs/superpowers/plans/2026-06-08-v2v-vertical-slice-forest-right-variants.md`. Tests: +4 (curve). Suite 228 passed / 2 skipped. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
benstull added 3 commits 2026-06-08 13:58:47 +00:00
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Right strength curve (level 1-4 -> increasing img2img keyframe strength) is pure
and unit-tested; the restyle engine is a verbatim port of the session-0008 POC
flow_restyle.py (SD img2img keyframes + Farneback optical-flow tweens = calm,
no boiling). torch/diffusers imported lazily so the module + curve test stay light.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Ran the baker for the forest scale: real flow-stabilized painterly Right variants
1-4 now back the dreamlike axis on real Yosemite footage (was: 1 real + 3 ffmpeg
placeholders). By-eye finding this surfaced: sd-turbo's painterly transform is
sharply non-linear in img2img strength (barely registers <0.5, ramps hard through
~0.5-0.65), so the initial linear 0.28-0.58 ramp left levels 1-3 ~identical to
raw. Re-baked with a clustered 0.46/0.52/0.58/0.64 curve -> a real perceptual
ramp: gentle dream haze -> clearly impressionist -> strongest, each notch visibly
dreamier, temporally calm (flow-stabilized).

- manifest: forest right_variants 1-4 all model 'sd-turbo+farneback-flow'
- setup_sample_media.py: now only stages the neutral base; the baker owns the
  Right variants, so re-running it never clobbers a real bake
- docs: USER_GUIDE setup steps + ROADMAP v2v vertical-slice-done note

Suite 228 passed / 2 skipped.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
benstull merged commit eace3e3467 into main 2026-06-08 13:58:54 +00:00
benstull deleted branch feature/v2v-forest-right-variants 2026-06-08 13:58:54 +00:00
Sign in to join this conversation.