Cavefishlabs/echodepth-video-analysis

feat: every memorability channel is reported as a place out of 100, and a prediction is compared with the median prediction

fix: Whisper medium with the brand as a hint, so a sung brand name at the end is heard; brand match tolerant of spelling

fix: each memorability channel's penalty is chosen under brand-grouped folds, so an unfamiliar name no longer lands anywhere

feat: memorability v0.2 reads the words and the brand's name, and can rescore a rewritten script against the same pictures

fix: CLIP image features spelled out, so transformers 5 on the Space reads them too

feat: memorability v0.1 model fitted to all of LAMBDA (held-out Spearman 0.52)

feat: memorability pass, predicted brand recall a day later from the pictures (model fitted to LAMBDA)

feat: the face pass returns each face's box and reading, and the frame size

feat: /proxy, a playable copy of a file without an analysis

feat: YuNet behind the four-frames-a-second emotion pass

fix: the proxy really does carry video and audio only

fix: proxy carries video and audio only, not the master's timecode track

feat: a browser-playable H.264 proxy for a master the browser cannot decode

fix: measure a film with no speech instead of returning nothing

fix: re-encode a film OpenCV cannot decode, instead of measuring nothing

feat: stills strip for the report; pin the face stack to the CEI v0.5 corpus environment

feat: Creative Effectiveness Index pass, screen text, brand check; pin OpenCV below 5

fix: install libEGL/libGLES so MediaPipe's face landmarker can init

fix: use post_init() so newer transformers registers tied-weight keys

feat: Stage Read pipeline and /stage-read routes

fix: normalize emotion2vec labels to clean English

feat: aggregate emotions from all detected faces

perf: increase emotion sampling to 4fps

perf: increase face emotion sampling from 0.5fps to 2fps

fix: correct HSEmotion-ONNX emotion label order

fix: switch to hsemotion-onnx for better portability

fix: use correct HuggingFace model paths for emotion2vec

chore: trigger rebuild to clear cached model failures

fix: use HuggingFace hub for emotion2vec download

fix: add better error logging for HSEmotion, try fallback model

fix: add missing dependencies for HSEmotion and emotion2vec

chore: force rebuild to update models endpoint

debug: add /models endpoint to check model loading status

feat: add HSEmotion + emotion2vec for real emotion detection

refactor: minimal v4.0 - stripped to essentials (whisper + wav2vec2 + opencv)

revert: restore bc0dc88 app.py without keep-alive changes

fix: add keep-alive progress during voice analysis to prevent SSE timeout

revert: restore midnight working state (bc0dc88) - 0.5 FPS hardcoded, no fps param

revert: restore 11pm working state (before fps/ffmpeg changes)

fix: add progress update during merge_timelines to prevent timeout

fix: use fps parameter from API request instead of hardcoded 0.5

fix: correct indentation in face analysis function

feat: use ffmpeg frame extraction instead of re-encoding

fix: simplify re-encoding to avoid subprocess deadlock

fix: skip NVENC, use ultrafast CPU with progress logging

perf: use NVIDIA NVENC for fast H.264 re-encoding

feat: add H.264 re-encoding for AV1/VP9 codec compatibility

feat: add keep-alive progress during vocal emotion analysis

chore: force Docker rebuild

feat: add /analyze_image endpoint for single image analysis