license: other
task_categories:
- video-text-to-text
tags:
- minimax-h3
- codex
- animation
- motion-graphics
- agentic-film-lab
pretty_name: Codex + MiniMax H3 Motion Graphics Lab — Evidence Record
Code animation over existing MiniMax H3 video
Evidence companion for Agentic Film Lab's 17:04.333 episode. The experiment adds editable, code-created graphics and animation to existing H3 footage, then examines separate H3 generations asked to create similar graphics inside the video.
Download codex-h3-motion-graphics-evidence-v001.zip for the complete staged record. The archive's FILE-MANIFEST.json records each file's public SHA-256 and the original hash where local-path sanitization changed a text file. MP4 evidence and synthesis stems are copied byte-for-byte. archive-sha256.json records the ZIP checksum.
Included
- All fourteen complete historical H3 results used in the episode: four initial field-card cells, six fresh-scene cells and four later replication cells, with native audio.
- The complete code demonstration sources identified in the delivery source inventory, the revised tape/keyboard HUD derivative, and the original freeze attempt alongside its repaired version.
- First-frame inputs, frozen prompts/seeds, API workflows, graphics code/configuration, sound synthesis stems, tracking geometry, mask-repair records, run ledgers, failed-attempt records, observations, timings and episode cost summary.
- Astra/Sol code comparison records, including the retained Max fireworks comparison. The rest of the final showcase uses Astra-authored code; parent integration, media handling and QC are separately recorded.
- Final narration text, captions, edit timeline and delivery QC. The creator's voice recording and narration audio are excluded.
How to read the record
The top-level historical drafts and earlier author folders document development; narration-read-v011.md and youtube-production/full-review-v005-tape-keyboard/ identify the final review. Original clips remain complete. Presentation zooms, labels, recaps, red grading, callouts and cartoon interactions are deterministic postproduction, not H3 object recognition or model-native text. H3-only replication regenerates the whole scene from a first-frame anchor; this is a workflow comparison, not a pixel-matched model ablation. Code targets were revised without new H3 generation.
The original freeze cutout missed part of the face. Its repair demonstrates the idea but remains imperfect. Other development corrections and technical failures remain in their recorded ledgers. Earlier portrait layouts are not used in this landscape episode. The copied tape-generator input contains superseded policy prose; the generator and audio-construction.json document the actual finite tape-reel and keyboard recipe.
Routes, timing and audio
MiniMax H3 ran as open weights through ComfyUI on RunPod; exact checkpoint hashes, workflows and hardware are in the ledgers. Generated comparisons have 243 frames at 24 fps (10.125 seconds), with the native complete audio preserved. Sound comparisons use the same moving HUD excerpt, source seconds 2–5. New code sound is finite tape transport/reel movement and physical key contacts; solo and combined examples use the same effects gain.
The final 4K episode is an upscale of the assembled 1080p edit; it does not establish native 4K H3 generation. Narration uses the creator's authorized voice via local Chatterbox, with spoken-script revisions authored by GPT-6 Astra. Cost summaries distinguish recorded RunPod lifecycle billing from unpriced model/subscription usage and local synthesis.
Scope and exclusions
This is a bounded production record, not a universal capability or quality ranking. Model weights and latent tensors, private voice-reference/narration audio, credentials, SSH keys, raw account billing responses, private Codex session history, redundant full-episode masters and decoded-frame caches are excluded. Text copies normalize private local path prefixes; some scripts still require adapting paths/dependencies to your own environment. Excluded large caches can be reconstructed from the included complete source clips where the code supports that process; this is an evidence archive, not a turnkey installation.
MiniMax model/media terms and other upstream licenses remain applicable; the other marker does not relicense them. Movie-reference code transcription and added comments are an illustrative Terminator-style reconstruction, not model-source code. No movie video is redistributed.
Sources
YouTube publication is pending; this repository is the public evidence destination.