Prophetero Test Bench + Take Compare GLOBAL
Test. Compare. Listen. Choose.
One workflow, several LoRA combinations—and a clearer way to choose the result worth finishing. Prepare short video trials, compare them on the same page, inspect the soundtrack, then carry the winning settings into a full-length render.
## Build experiments without juggling tabs
- Capture your working ComfyUI video workflow; create a LoRA-free baseline and independent sample cards.
- Search installed LoRAs and set each card's switches and strengths. Every card shows its own saved active LoRA names and values.
- Use automatic short trials for recognized H3/LTX timing layouts, choose a custom length, or keep the original duration.
- Run prepared samples sequentially. Add a third or fourth experiment later: completed takes stay available and are not regenerated.
- Start controls sit directly below the cards. Stop after the current test or cancel this test's own job; keep completed results.
## Compare picture, motion and sound
Take Compare is built into the test page. Review up to four clips with shared playback, frame stepping, linked zoom, timing offsets, notes and marked moments. Each video also has independent play/pause, 0.25x–2x speed and repeat, so you can inspect one take at a time on a single monitor.
Finished sample cards have their own players. Live image frames can appear during generation when the workflow sends supported standard ComfyUI previews.
Open Audio inspection to load the original soundtrack's mono/stereo waveforms, zoom and seek, and inspect decoded sample peak/RMS levels in dBFS. Choose picture, motion and audio winners separately.
## Keep the winner
Prepare a full-length job from a completed sample, then explicitly start it. Or export the winning workflow with its saved settings and restored full duration. Verified mappings export editable UI JSON; dynamic or unverified mappings use a clearly labelled API JSON fallback.
Not just H3. The reviewer accepts supported local videos from LTX, other AI generators, cameras and editors. Generation testing requires a working ComfyUI graph with supported loaders/savers and your installed resources. Optional handoffs to Take Mixer, Loop Lab and Final Check are available when those separate tools are installed.
## Install and start
Extract the entire ZIP. Stop your backend, run START_HERE.cmd, and choose the ComfyUI backend your application actually uses. Restart the entire SwarmUI/ComfyUI backend and refresh the browser. Open a working generation workflow and right-click empty canvas -> Prophetero · Test Bench + Take Compare.
Windows installer; English default; EN / HU / DE / FR / ES / IT / PT / PL. Models, LoRAs, Python/ComfyUI, other companion tools and personal videos are not bundled. No paid cloud account is required.
## Before downloading
Public beta 0.4. Supported outputs: SaveVideo and VHS_VideoCombine. LoRA editing supports Power Lora Loader (rgthree), LoraLoader and LoraLoaderModelOnly. Model selection does not convert architectures or fix encoder/VAE compatibility. Automatic shortening depends on recognized timing controls; a short successful test cannot guarantee a full-length render.
Review limits: 0.5–120 seconds, 1–60 FPS, up to 8 megapixels and 2 GiB per clip / 6 GiB total; upright square-pixel SDR video and mono/stereo audio. Up to 24 samples/project, four compared at once. Playback proxies leave originals unchanged; source frames are used for paused detail. Waveforms are sample peak/RMS measurements, not LUFS/true-peak mastering or automatic repair. Shared playback is a review aid, not guaranteed frame/sample-locked mastering. Visual/anatomical defects still need your judgement.
Isolated browser, audio, workflow-planning and installer tests passed using neutral clips and simulated inference. Real-model coverage remains limited; see the included validation notes.
Credits: built for ComfyUI and the user's existing generation ecosystem. Thanks to the ComfyUI, FFmpeg, PyAV, imageio-ffmpeg, Pillow, NumPy, rgthree and VideoHelperSuite contributors. Model/LoRA creators retain credit and their own terms; see THIRD_PARTY.md.
Usage: free personal and commercial creative work, with free sharing/modification under the included licence. Selling this software or modified versions is not permitted. Retain its licence/notices when sharing. Your source-media and generation-model rights remain separate.
If it helps you find a better take—or saves you a few unnecessary renders—I'd love to hear about your setup and what you would improve next.
Description
# Prophetero Test Bench 0.4 + Take Compare 0.2
## Public beta — November 5, 2026
- Build a LoRA-free baseline and independent sample cards with searchable LoRAs, per-card switches/strengths and visible saved settings.
- Run prepared trials sequentially, using supported automatic short-duration profiles, custom timing or original full length.
- Keep completed takes when adding new experiments; Start runs only unstarted prepared samples.
- Run controls directly below sample cards, with prepared/completed counts, stop-after-current and targeted cancellation.
- Play finished videos within their cards; show live images when the source workflow emits supported standard ComfyUI preview frames.
- Integrated Take Compare: review up to four takes on the same page, with frame stepping, timing offsets, linked zoom, notes and markers.
- Shared selected-range repeat and 0.25x–2x viewing speed; independent play/pause, speed and repeat controls on every comparison video.
- Optional original-source audio waveforms, separate mono/stereo channels, zoom, seeking, sample peak/RMS and time-window levels.
- Preserve unfinished edits locally, block accidental runs with unsaved drafts, remove mistaken plans without deleting video files.
- Prepare a full-duration render from a finished sample, or export its exact winner settings using verified UI JSON / labelled API JSON fallback.
- Eight interface languages: EN, HU, DE, FR, ES, IT, PT, PL; English-first public package and generic Windows installer.
### Release notes
Public beta. Tests covered isolated browser use, fixture audio, workflow planning, preview protocols and installer behaviour. Inference was simulated; universal real-model/custom-node compatibility is not claimed. Waveform metrics are decoded sample peak/RMS, not LUFS or true peak. Preview/playback adjustments leave source media unchanged. Detailed test scope is included in VALIDATION_TEST_BENCH.md.
