The Runbook

e studios remix — Session Handoff / Next-Session Runbook

Written 2026-07-08 at the close of the founding session. Read this + E_STUDIOS_REMIX_SPEC.md first.

Start future sessions with cwd = ~\Downloads\e_render or C:\AI.

Current state (all verified)

• Three finished cuts in e_render: e_studios_remix_v1_full.mp4 (2023 print), _v1_1 (fade coda), _v1_2 (combustion coda). Original audio remuxed on all.
• Parts library: 56 parts in C:\AI\ComfyUI_windows_portable\ComfyUI\output (RUNTMP/CODA/PEANUTS/IPA prefixes). Concat lists in e_render\v1full\concat*.txt.
• Backup: full structured archive at iCloudDrive\e_studios_remix (270 files, 571MB).
• Stack: ComfyUI portable @ C:\AI\ComfyUI_windows_portable, torch 2.12.1+cu126, nodes: AnimateDiff-Evolved, VideoHelperSuite, Advanced-ControlNet, FizzNodes, IPAdapter_plus. Server: python_embeded\python.exe -s ComfyUI\main.py --port 8188 --disable-auto-launch. Drive via API (/object_info, /prompt, /history).
• Scripts (C:\AI): render_shots.py (68-shot manifest + per-run renderer — canonical), render_38.py (full v1 run + 38-style list), coda_travel*.py (prompt travel; v12 = + CN decay schedule), ipa_recrank.py (generic IP-Adapter re-crank: canny_dir, ref, out, prompt, [CN], [IPW]).
• Models on disk: SD1.5 fp16, mm_sd_v15_v2, canny CN fp16, ip-adapter-plus_sd15 + CLIP-ViT-H.
• Docs: spec, artist statement draft, shot bible (e_shots\INVENTORY.md), prototype index (proto_index.png; 57 refs in e_render\aiimgs_tmp_iCloud\iCloud Photos; #12=comic strip, #33=ligne claire Palantir). NB: iCloud filesystem order ≠ proto_index numbering — the ligne claire Palantir is staged as ComfyUI\input\ligne_ref.png (the seg09 IPA ref); don't grab prototypes by nth filename.

Doctrines learned (don't relearn)

1. Per-shot content prompts, style templates, bell-strike transitions. Edges carry lettering at moderate style distance; brand color dies first, wordmark last.
2. CN strength by style family: photographic ~0.9 / painterly 0.85 / high-abstraction 0.5–0.65 + early end_percent. Cubism categorically impossible under edge-lock (needs Kontext/Qwen). Line styles need lineart CN (not yet downloaded), not canny+IPA.
3. IP-Adapter transfers flatness+palette, not pen line. Always state palette in text prompt. Recipe: ref image = pen, prompt = palette, CN = permission.
4. Prompt travel (BatchPromptSchedule) works; for full dissolution also schedule CN via LatentKeyframeTiming→TimestepKeyframe (v1.2 proof).
5. Shot table drifts ~0.3–0.5s late in the 3.0s and 6.4–8.0s+ neighborhoods — fix boundaries before precision work, BUT all four drift accidents outperformed intent (ICE→constructivist, kick→clay, Grindr→sprites, DreamOn→psychedelic gatefold). Canonize accidents in cut files.
6. VRAM: 6GB = 512² @ 77+ frames OK; 720² OOM; never batch Canny on GPU (precompute CPU, OpenCV 64/166).

Next-session runbook (in order)

1. Driver update DONE — 610.62. Everything unlocked.
2. Lineart CN DONE — ligneclaire_lineart_v3.mp4. Verdict: the pen arrived (line ~90%) — doctrine confirmed: line styles need the lineart conduit (aux pack now installed; LineArtPreprocessor needs coarse='disable' passed explicitly). Remaining gap: palette ~50% (sepia only). v4 DONE — ligneclaire_lineart_v4.mp4 (lignev4.py): named colors in POS delivered the palette (~85% — blue body, brick red accents, cream ground; stable across frames). Recipe closed: ref=pen, prompt=palette, lineart CN=permission. Five-gen comparison: progress_ligne_5gen.png.
3. Local taste tests DONE — both. LTX 2B distilled (ltx2b_s24_taste.mp4): 97f/1min, drifts off-composition, filmic melt — cast as transition tissue + coda texture, not segment renderer. CogVideoX-2B (cog2b_s24_taste.mp4): from text alone dreamed a stable NYSE balcony ceremony with glossolalia frieze lettering — Chinese-print thesis confirmed; 14min/clip parts oven. Scripts: ltx_taste.py, cog_taste.py.
4. Local LoRA DONE — trained (diffusers not kohya; 2000 steps, rank 16, 15min not hours). Weights: C:\AI\lora_train\out\pytorch_lora_weights.safetensors (also in ComfyUI loras as estudios_remix_style.safetensors). Audition lora_audition_grid.png: coherent inky editorial-comic house style, saturates by w=0.6 → use 0.3–0.4 as glue; not yet tested with CN on real shots.
5. Cloud weekend (~$20 RunPod 4090): Wan 2.2 14B v2v all segments = "the Wan restoration"; Qwen-Image-Edit hero frames (ICE/Palantir/Grindr) + the cubist segment solve. TURNKEY: cloud/bootstrap.sh (verified model files) + cloud/CLOUD_SESSION.md (render plan → maximal version). Base cut for cloud swaps = cuts/v1_cast.
6. "The Chinese print" — full alternate version rendered end-to-end with Chinese-origin models (Wan/Qwen/Kolors): the American mania re-rendered by the successor's engines. Release as its own version, per the versions-as-media-history doctrine.
7. Librarian + assemble.py DONE 2026-07-08. C:\AI\librarian.py recovers per-part provenance from ComfyUI's embedded workflow metadata (segment/style/model/CN/seed/frames) → catalog.json (63 parts, all workflow-recovered). C:\AI\assemble.py builds finished mp4s from JSON EDLs in cuts\. Verified: rebuilds shipped master exactly (40.099s), and regenerates the v1 EDL from metadata alone (concat.txt now redundant). Usage: LIBRARIAN.md.

Standing priorities above all of it

• TRANSMIT v1 first. Statement drafted, congregation waiting, Contain return-episode is the natural premiere vector. Restorations come after release.
• The film satisfies the monthly-artifact rule only if posted.