5982801cae
The endorsed design: the I-frame is the LEARNED integer image codec int_hyper (wai.neural.int_hyper), decoded byte-exact, and the P-frames are integer block-MC residuals on top — so the whole video decode is integer and byte-identical on every machine. A learned keyframe you can put in a receipt, plus integer inter-frames. - int_video::decode_pframes(i_frame, h, w, bs): the P-frame loop seeded with ANY decoded I-frame (integer MC from the previous reconstructed frame + quantized residual over rANS). Decouples the inter-frame codec from the keyframe source. - The NEURAL-KEYFRAME path = int_hyper_synth::decode_conditional (frame 0) ∘ decode_pframes (the rest). Both pure i64; the composition lives in the test/caller. - tools/wai_int_video_keyframe_encode.py: uses the committed tiny int_hyper vector as frame 0 (so the conformance is self-contained, never-skips), translates it for the P-frames, codes the residual. Emits only the P-frame bitstream + expected hashes (frame 0's hash IS int_hyper's, 7cbe882c). - Committed vector (tests/vectors/int_video/video_nk*) + NEVER-skip test (gated on neural_int+video_int): decodes the I-frame via int_hyper then the P-frames, every frame byte-for-byte + clip hash a5f53d477999d159; frame 0 == the int_hyper image. - Wired wai.video.int_hyper: NeuralIntegerExact (embeds the learned integer codec), Lossy tier. `wai capabilities` → 58, 3 integer-exact (int_synth, int_hyper, video). SPEC table + capability comment. Co-Authored-By: Claude Opus 4.8 <[email protected]>
1 line
No EOL
184 B
JSON
1 line
No EOL
184 B
JSON
{"n_frames": 5, "h": 16, "w": 16, "frame_hashes": ["7cbe882c42c6725d", "51d72863061a7af8", "7f212e2c709104f4", "18b67d5614767fb1", "8ff3a77836a2f19a"], "clip_hash": "a5f53d477999d159"} |