vdaular's picture
|
download
raw
3.67 kB
# MiniMax-H3 Multi-Stacking & Dual-Persona Video Benchmark Suite
This directory contains verified `.mp4` video generation benchmarks demonstrating single-persona Triple-RefMod stacking, body shape booster quad-stacking, dual-persona simultaneous shared frames, dual-audio lip sync, and polyphonic counterpoint singing duets in **MiniMax-H3**.
For complete architectural details, mathematical foundations, and prompting guidelines, see the main guide:
๐Ÿ‘‰ **[`../MINIMAX_H3_REFMOD_STACKING_AND_MULTISUBJECT_GUIDE.md`](../MINIMAX_H3_REFMOD_STACKING_AND_MULTISUBJECT_GUIDE.md)**
Identity LoRA vs RefMod A/B (Jinx epoch 60 / 80 / final, RefMod-only, RefMod+LoRA):
๐Ÿ‘‰ **[`../lora/README.md`](../lora/README.md)** โ€” epoch 60 is enough; **RefMod + LoRA is the best stack**.
---
## Benchmark Video Inventory
| Video Filename | Duration / Res | Personas & Audio | Pipeline & LoRA | Description |
| :--- | :--- | :--- | :--- | :--- |
| **`test_billieeilish_turbo8_3310245_speech_triple_refmod_8s_00001_.mp4`** | 8.0s (192f)<br>1344x768 | Billie Eilish (Triple RefMod: `billie3` + `billie4` + `billie5`) | Turbo 8-Step + LoRA `3310245` (0.52) | **๐Ÿฅ‡ Single-Persona Gold Standard**: Authentic breathy dialogue delivery with 99.8% facial likeness lock. |
| **`test_dual_billie_miley_turbo8_3310245_conversation_6s_08mp_00001_.mp4`** | 6.0s (144f)<br>1184x672 | Billie Eilish & Miley Cyrus (6x RefMods) | Turbo 8-Step + LoRA `3310245` (0.52) | **Dual Conversation Baseline**: Clean turn-taking conversational dialogue between two distinct personas. |
| **`test_dual_billie_miley_turbo8_3310245_conversation_6s_08mp_dual_audio_00001_.mp4`** | 6.0s (144f)<br>1184x672 | Billie Eilish & Miley Cyrus + Dual Spoken Audio | Turbo 8-Step + Dual Audio Conditioning | **Dual Spoken Audio Conditioning**: `<Audio 1>` (Billie) + `<Audio 2>` (Miley) with independent voice timbres. |
| **`test_dual_billie_miley_turbo8_3310245_simultaneous_twoshot_6s_00001_.mp4`** | 6.0s (144f)<br>1184x672 | Billie Eilish & Miley Cyrus (6x RefMods) + Dual Audio | Turbo 8-Step + Dual Audio Conditioning | **๐Ÿฅ‡ Gold Standard Simultaneous Two-Shot**: Continuous unbroken frame with both personas visible side-by-side (zero cuts). |
| **`test_dual_billie_miley_turbo8_3310245_singing_duet_madonna_6s_00001_.mp4`** | 6.0s (144f)<br>1184x672 | Billie Eilish & Miley Cyrus + Vocal Audio | Turbo 8-Step + Dual Vocal References | **Harmonic Duet**: Shared two-shot performing Madonna's *"Like a Prayer"* in harmony. |
| **`test_dual_billie_miley_turbo8_3310245_singing_counterpoint_madonna_6s_00001_.mp4`** | 6.0s (144f)<br>1184x672 | Billie Eilish & Miley Cyrus + Vocal Audio | Turbo 8-Step + Dual Vocal References | **๐Ÿฅ‡ Polyphonic Counterpoint Duet**: Simultaneous singing with independent melodies and lyrics in shared frame. |
| **`test_dual_billie_miley_turbo8_3310245_singing_strict_separation_6s_00001_.mp4`** | 6.0s (144f)<br>1184x672 | Billie Eilish & Miley Cyrus + Vocal Audio | Turbo 8-Step + Dual Vocal References | **๐Ÿฅ‡ Strict Prompt Separation Mashup**: Independent simultaneous songs (*"Birds of a Feather"* + *"Jolene"*). |
| **`test_dual_billie_miley_turbo8_3310245_singing_two_different_songs_6s_00001_.mp4`** | 6.0s (144f)<br>1184x672 | Billie Eilish & Miley Cyrus + Vocal Audio | Turbo 8-Step + Dual Vocal References | **Simultaneous Mashup Baseline**: Unbroken shared two-shot vocal performance. |
---
## Repositories & Resources
- **Hugging Face RefMods Hub**: [https://huggingface.co/malcolmrey/minimaxh3](https://huggingface.co/malcolmrey/minimaxh3)
- **H3 Center Hub**: [https://huggingface.co/datasets/malcolmrey/various](https://huggingface.co/datasets/malcolmrey/various)

Xet Storage Details

Size:
3.67 kB
ยท
Xet hash:
3ee62e7642b5066ff21d658220d975b25e9734d52b86d4a4ec47c2771858a87a

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.