examples

h3-center/docs/C2V_CONTINUOUS_CHAIN_GUIDE.md

@@ -224,4 +224,110 @@ Do **not** copy the Miley Cyrus full-day walk (terrace → kitchen → living
 4. After clip 1, A/B LoRA +0.05 on that seed only. Natalia: 0.45/0.45 (`01962`) beat 0.40/0.40 (`01957`).
 5. Only then continue the chain or start a longer day series.
 
+---
+
+## 8. Inter-Scene Bridging: C2V vs True FL2VA (First/Last Frame Keyframe Interpolation)
+
+When bridging two separate, already generated scenes (e.g. Scene 1 ending on an environmental wreckage $\to$ Scene 2 opening on an arriving character), standard C2V with R2V prompt text is insufficient: C2V carries motion forward from Scene 1, but cannot mathematically constrain the model to land precisely on Scene 2's starting composition, posture, and lighting.
+
+### The True FL2VA Bridge Architecture
+
+```text
+[Scene 1 MP4] --(extract tail frame)--> [scene01_last_frame.jpg]  (first_frame)
+                                                                 \
+                                                                  --> [MiniMaxH3ImageToVideo] --> [Bridge Clip]
+                                                                 /
+[Scene 2 MP4] --(extract head frame)--> [scene02_first_frame.jpg] (last_frame)
+```
+
+1. **Extract Boundary Frames with FFmpeg:**
+   ```bash
+   # Scene 1 tail frame (first_frame)
+   ffmpeg -y -sseof -0.05 -i movie_scene01.mp4 -vframes 1 -q:v 1 scene01_last_frame.jpg
+
+   # Scene 2 head frame (last_frame)
+   ffmpeg -y -ss 00:00:00.000 -i movie_scene02.mp4 -vframes 1 -q:v 1 scene02_first_frame.jpg
+   ```
+
+2. **Dual-Keyframe Conditioning:**
+   - Link `first_frame` and `last_frame` to `MiniMaxH3ImageToVideo` (or `MiniMaxH3FLConstraint`).
+   - The DiT pins frame 0 to `scene01_last_frame.jpg` and frame $T-1$ to `scene02_first_frame.jpg`, solving a physically smooth, continuous camera move and motion arc between the two keyframes.
+
+3. **Important Architectural Rule: No RefMod on FL2VA Dual Keyframes:**
+   - Applying `MiniMaxH3RefModApply` on top of dual-keyframe conditioning causes a tensor shape mismatch in the sampler (`RuntimeError: shape mismatch [5888, 96] vs [7904, 96]`).
+   - Because both the head and tail frames are already hard-pinned, subject likeness, wardrobe, and environment textures are completely locked by the image latents without needing extra adapter injection.
+
+4. **Lossless Video Concatenation (`ffmpeg -c copy`):**
+   ```bash
+   # scenes_list.txt:
+   # file 'movie_scene01.mp4'
+   # file 'movie_scene01b_bridge.mp4'
+   # file 'movie_scene02.mp4'
+   ffmpeg -y -f concat -safe 0 -i scenes_list.txt -c copy full_movie_stitched.mp4
+   ```
+   - Transitions have **zero visible seams, 0ms audio/visual jumps, and 100% continuous flow**.
+
+---
+
+## 9. Key Production Directives & Audio/Visual Artifact Prevention (September 2026)
+
+### 1. Motion Context Trim & Dialogue Lead-In Mechanics
+- **The Trim Offset Phenomenon**: MiniMax-H3's internal TTS engine synthesizes audio starting at `00:00.000` of the *untrimmed* timeline. When `MiniMaxH3MotionContextTrim` cuts the first $N$ context frames (e.g. 22 frames = ~0.92s), the corresponding head audio is sliced away. If the character begins speaking immediately at frame 0, the first 1–2 words (e.g. `"So we already..."` $\to$ `"already..."`) are lost.
+- **Ineffectiveness of Prompt Timecodes for TTS Delay**: Writing explicit timing directives (e.g. `From 00:00.000 to 00:01.500, he pauses silently...`) guides video diffusion but is ignored by the TTS encoder, which generates speech from the very beginning.
+- **The Overlap Carryover Tag Solution**: Prepend the trailing 1–2 words of the *previous* clip's spoken dialogue into the current clip's `<d>[English] ...</d>` tag.
+  - *Example*: If Clip \(N-1\) ended with `"...running on our stack."`, begin Clip \(N\)'s prompt dialogue with:
+    `<d>[English] our stack. So we already produce Banco de España reporting...</d>`
+  - The prepended words `our stack.` land inside the 22-frame trim window and are cut, allowing the new dialogue sentence to begin cleanly at `00:00.000` of the delivered clip.
+
+### 2. Dialogue Pacing vs Clip Duration (Preventing "Talking to Himself" & Filler Hallucination)
+- **The Hallucination Trigger**: MiniMax-H3 attempts to fill the entire generated duration with audio. If a short sentence (5–7 seconds of speech) is allocated to a 15-second clip, the model synthesizes the requested sentence and then **hallucinates unscripted filler speech**, repetitive loops (`"You You You..."`), or secondary dialogue where the character appears to talk to himself or someone off-camera.
+- **Calibration Formula**:
+  $$\text{Spoken Duration (s)} \approx \frac{\text{Word Count}}{2.5 \text{ to } 3.0 \text{ words/sec}}$$
+  $$\text{Total Generation Frames} = \text{round}\big((\text{Spoken Duration} + 0.8\text{s buffer}) \times 24\big) + \text{trim\_frames (22)}$$
+- Always calibrate clip duration tightly to sentence length rather than enforcing arbitrary fixed 15-second blocks.
+
+### 3. Motion Context Sizing Calibration: 22 Frames vs 56 Frames
+- **22 Frames (0.92s)**: The optimal sweet spot for unbroken dialogue and subtle physical motion. Produces flawless visual seams ($\le 3.5$ mean pixel difference) with zero headstart warping.
+- **56 Frames (2.33s)**: Use only for large physical locomotion (e.g., walking through doors). In static monologues, 56 frames can introduce headstart jitter and requires a significantly larger audio overlap.
+
+### 4. Resolution Calibration for Fast Production
+- **0.4 Megapixels (832x480)**: Outstanding rendering speed (~75–80s per 8–9s clip on 8-step Turbo) while preserving complete biometric identity and crisp facial expressions with RefMods.
+
+### 5. Progressive Quality Degradation & The Shot-Cut Strategy
+- **The Latent Drift Accumulation**: While C2V preserves continuous motion across adjacent clips, chaining 5+ consecutive clips in a single unbroken locked camera angle causes progressive latent drift, slight loss in facial crispness, and subtle rendering artifacts.
+- **The Cinematic Shot-Cut Solution**: In multi-minute monologues and feature scenes, introduce **logical camera angle changes / cinematic cuts** every 3–4 clips (30–45 seconds).
+  - *Cut Types*: Switch between Medium Shot $\to$ Over-the-shoulder $\to$ 3/4 Profile $\to$ Close-up $\to$ Cutaway to documents/props $\to$ Return to Master MCU.
+  - *Benefits*: Every cut allows seeding a fresh clip with clean latent conditioning, completely resetting quality and sharpness to 100% while improving cinematic pacing and visual interest.
+
+### 6. Two-Tier Resolution Strategy: Low-Res Testing vs High-Res Commits
+- **0.4–0.5 Megapixels (480p / 544p)**: Use heavily during development to rapidly iterate on prompt composition, movement mechanics, dialogue duration calibration (testing 10s vs 11s vs 12s), and Whisper audio verification.
+- **~2.0 Megapixels / Full HD (1080p / 1344x768)**: Once prompt, motion, and dialogue timing are verified, commit hero dialogue close-ups and key scenes at ~2 MP. Full HD eliminates the plastic "AI video" sheen, producing realistic skin pores, micro-expressions, and believable eye contact.
+
+### 7. Explicit 5-Parameter Static Camera Lock
+- When a static camera must not drift over a long dialogue take, dictate all 5 spatial parameters explicitly in prompt:
+  1. `camera distance`: (e.g. `medium close-up, chest-up distance`)
+  2. `character scale`: (e.g. `character scale strictly maintained`)
+  3. `framing`: (e.g. `subject seated slightly off-centre left`)
+  4. `background position`: (e.g. `background perspective remains completely stationary`)
+  5. `character position`: (e.g. `remains seated in the exact same position`)
+- *Mandate*: `Completely fixed, locked-off static camera position throughout the entire video. The camera remains completely stationary, and the framing, character scale, and background position remain strictly matched to <Picture 1> with zero camera movement, no panning, no tilting, no zooming, no push-in, and no recomposition.`
+
+### 8. Strict Isolation of Off-Screen Cast & Dialogue Separation (J/L-Cuts)
+- **Never Name Off-Screen Characters**: In single-character close-ups or reactions, never mention off-screen characters in descriptive prose (e.g., avoid "Data reacts to Picard"). Describe solely what the on-screen subject does to prevent facial blending or phantom character generation.
+- **Off-Screen Dialogue Separation**: If Character A speaks while looking at Character B:
+  1. Generate a clean close-up of Character A speaking the line on-camera (pristine audio).
+  2. Generate a clean close-up of Character B reacting in multi-layered silence.
+  3. In NLE / FFmpeg, extend Character A's audio across the cut into Character B's reaction shot (J-Cut / L-Cut).
+
+### 9. 4-Point Multi-Layered Silence Protocol for Reaction Shots
+- Simply writing `"no dialogue"` fails if the character's mouth is slightly parted in the reference image.
+- *Mandate*: (1) slightly parted mouth is strictly a resting facial posture and not speech, (2) lips do not form words, (3) jaw makes no speaking movements, (4) character does not mouth dialogue and remains completely silent.
+
+### 10. Inline Performance Tags & Speech Controls
+- Use inline tags inside `<d>[Language] ...</d>` for nuanced human speech delivery:
+  - Natural pauses & breathing: `<pause>`, `<long pause>`, `<breath>`, `<inhale>`, `<exhale>`, `<deep breath>`, `<sighs>`
+  - Vocal tone & delivery: `<whisper>`, `<softer>`, `<stutter>`, `<laughs>`, `<chuckle>`
+  - Word emphasis: `<i>word</i>` (e.g. `<d>[English] I did <i>not</i> agree to this.</d>`)
+
+
 

h3-center/docs/MINIMAX_H3_PROMPTING_GUIDE.md

@@ -528,6 +528,128 @@ This mirrors high-caliber cinema pacing and gives each MiniMax-H3 latent step op
 
 **The rule:** Refs and LoRA strengths lock from the **same** winning clip. Do not mix refs from clip A with strengths from clip B without a re-test.
 
+### Rule 46: True Bridge Scene Creation via Dual-Keyframe FL2VA & Boundary Frame Extraction (September 2026)
+**The Problem / Failure:**
+When connecting two independently generated scenes (e.g. Scene 1 ending on an environmental wreckage shot $\to$ Scene 2 opening on a character standing/arriving), using standard C2V (motion context from Scene 1) with generic R2V prompt text fails to create a seamless bridge. The model hallucinates an arbitrary trajectory, ignores the destination framing, or fails to terminate on the exact framing, composition, lighting, and posture of Scene 2's opening frame.
+
+**The Solution — True FL2VA Dual-Keyframe Bridge Architecture:**
+1. **Extract Exact Boundary Frames with `ffmpeg`:**
+   - **Scene A Tail Frame (`first_frame`):**
+     ```bash
+     ffmpeg -y -sseof -0.05 -i scene_01.mp4 -vframes 1 -q:v 1 scene01_last_frame.jpg
+     ```
+   - **Scene B Head Frame (`last_frame`):**
+     ```bash
+     ffmpeg -y -ss 00:00:00.000 -i scene_02.mp4 -vframes 1 -q:v 1 scene02_first_frame.jpg
+     ```
+2. **Dual-Keyframe Conditioning via `MiniMaxH3ImageToVideo`:**
+   - Feed `first_frame` (Scene A tail) and `last_frame` (Scene B head) directly into `MiniMaxH3ImageToVideo` (or `MiniMaxH3FLConstraint`).
+   - The H3 model's native FL2VA encoder mathematically pins frame 0 to `first_frame` and frame $T-1$ to `last_frame`, computing a smooth, physically grounded camera and motion trajectory between the exact boundary states.
+3. **Reference Framing Anchors the Scene Space:**
+   - In the prompt, describe how the transition unfolds across the timeline, starting from the exact environment of `first_frame` and fluidly settling into the exact framing and pose of `last_frame`.
+4. **Architectural Gotcha — RefMod Incompatibility with FL2VA:**
+   - **The Error:** Applying `MiniMaxH3RefModApply` on top of dual-keyframe FL2VA conditioning triggers a tensor dimension mismatch in the sampler:
+     `RuntimeError: shape mismatch: value tensor of shape [5888, 96] cannot be broadcast to indexing result of shape [7904, 96]` (due to the extra keyframe latent tokens modifying the attention sequence length).
+   - **The Fix:** For FL2VA bridge scenes, use pure `MiniMaxH3ImageToVideo` without `MiniMaxH3RefModApply`. Because both the start and end frames are hard-pinned, character identity, costume, face structure, and environmental lighting are already 100% physically locked by the keyframes.
+5. **Lossless Stream Stitching with FFmpeg (`-c copy`):**
+   - Because all clips share identical stream properties (1344x768 @ 24fps, H.264 + 32kHz AAC), concatenate with zero re-encoding:
+     ```bash
+     ffmpeg -y -f concat -safe 0 -i scenes_list.txt -c copy full_movie.mp4
+     ```
+   - The transition is completely invisible with 0ms visual jump, seam, or color shift.
+
+### Rule 47: Generation Duration Directly Governs Dialogue Cadence & Speech Rate (September 2026)
+**The Problem:**
+Allocating the wrong generation duration for a given line of dialogue causes severe performance distortions:
+- **Too short duration**: The actor races through lines unnaturally fast, slurring words to squeeze the entire sentence before the clip cuts.
+- **Too long duration**: The actor delivers lines unnaturally slowly and drawn out, or finishes early and fills remaining seconds with hallucinatory filler speech, repetitive loops, or mumbling.
+
+**The Test-and-Commit Workflow:**
+1. **Low-Resolution Duration Calibration (480p / 0.4 MP)**: For dialogue shots (especially 10–15 seconds), test candidate durations in 1-second increments (e.g. 10s, 11s, 12s, 13s) at low resolution.
+2. **Lock-in**: Find the exact duration where speech pacing, pauses, and physical cadence feel completely natural and human.
+3. **High-Resolution Commit**: Only commit to the expensive high-resolution / 2 MP generation once the exact duration and pacing are verified.
+
+### Rule 48: High-Resolution Generation (2 Megapixels / Full HD) Eliminates "AI Video" Artifacts (September 2026)
+**The Insight:**
+While 0.4–0.5 MP (e.g. 832x480 or 960x544) is ideal for rapid testing of movement, dialogue pacing, and framing, generating final hero shots around the **2-Megapixel / Full HD range** (e.g. 1920x1080 or 1344x768) fundamentally transforms character presence:
+- **Facial Fidelity**: Eliminates the smooth plastic "AI video" sheen, delivering genuine skin pores, subtle micro-expressions, realistic eye reflections, and believable dental/lip geometry.
+- **Production Standard**:
+  - *Low Resolution (0.4 MP)*: Rapid iteration, prompt verification, duration calibration, audio sync testing.
+  - *High Resolution (2.0 MP)*: Final production renders for dialogue close-ups, emotional reactions, and hero scenes.
+
+### Rule 49: Visual Authority of Starting Reference Images Over Massive Prompts (September 2026)
+**The Insight:**
+A pristine, well-composed starting image (or keyframe) is far more potent than multiple pages of text instructions.
+- If character placement, costume, set architecture, lighting temperature, camera angle, and lens focal length are already correct in the starting image, H3 has far less opportunity to drift or hallucinate.
+- **Principle**: Treat the reference image as the **visual authority** for the shot. Make the starting frame as close as possible to the desired look before generating video, rather than trying to steer an imperfect frame with bloated text prompts.
+
+### Rule 50: Explicit Static Camera Locking Vocabulary (September 2026)
+**The Problem:**
+Simply asking for a "static shot" or omitting camera motion often allows H3 to slowly push in, drift sideways, or subtly recompose the frame over a 10–15s clip.
+
+**The Solution — Complete Parameter Lock:**
+Explicitly dictate all five spatial parameters when a locked camera is required:
+1. `camera distance`: (e.g. `medium close-up, chest-up distance`)
+2. `character scale`: (e.g. `character scale strictly maintained`)
+3. `framing`: (e.g. `subject seated slightly off-centre left`)
+4. `background position`: (e.g. `background perspective and room elements remain completely stationary`)
+5. `character position`: (e.g. `remains seated in the exact same position`)
+- **Explicit Mandate**: `Completely fixed, locked-off static camera position throughout the entire video. The camera remains completely stationary, and the framing, character scale, and background position remain strictly matched to <Picture 1> with zero camera movement, no panning, no tilting, no zooming, no push-in, and no recomposition.`
+
+### Rule 51: Strict Isolation of Off-Screen Characters (Never Name Non-Visible Cast) (September 2026)
+**The Problem:**
+In single-character close-ups or reaction shots, referencing off-screen characters in descriptive prose (e.g., `"Data reacts to Picard with a subtle head tilt"` or `"Troi listens as Riker speaks"`) frequently confuses H3 into:
+- Spawning a second person or body part into the frame.
+- Morphing or blending the visible actor's facial features with the off-screen character.
+
+**The Solution:**
+- In isolated single-character shots, **never mention any character who is not physically visible on screen**.
+- Describe exclusively what the visible character is doing: `Data tilts his head slightly with a subtle, inquisitive analytical expression, listening intently off-camera.`
+
+### Rule 52: Off-Screen Dialogue & Reaction Coverage Separation (J-Cuts / L-Cuts) (September 2026)
+**The Problem:**
+Prompting a character to speak off-screen while the camera focuses on another character often results in H3 moving the visible character's mouth or corrupting the vocal track.
+
+**The Solution — Film Coverage Separation:**
+1. **Shot A (Dialogue Capture)**: Generate a clean close-up of the speaking character delivering the line on-camera to get pristine, uncorrupted audio.
+2. **Shot B (Reaction Capture)**: Generate a separate shot of the listening character reacting in complete multi-layered silence (Rule 53).
+3. **Editorial Assembly (NLE / FFmpeg)**: In Adobe Premiere Pro, DaVinci Resolve, or FFmpeg, create a **J-Cut or L-Cut** by extending the speaker's audio across the cut into the reaction shot. This matches traditional cinematic film coverage and feels vastly more natural than locked single-angle monologues.
+
+### Rule 53: Multi-Layered Silence Mandates for Reaction Shots (September 2026)
+**The Problem:**
+Simply writing `"no dialogue"` or `"remains silent"` is often ignored by H3 if the subject's mouth is slightly open in the reference image, leading to random mumbling or mouth flapping.
+
+**The Solution — 4-Point Anatomical Silence Protocol:**
+When generating silent reaction shots or listening coverage, explicitly specify:
+1. `The slightly open mouth is strictly a resting facial posture and not the beginning of speech.`
+2. `The lips do not form words or move.`
+3. `The jaw makes no speaking movements.`
+4. `The character does not mouth dialogue or utter any sound, remaining completely silent throughout.`
+
+### Rule 54: Inline Speech & Performance Control Tags (September 2026)
+**The Feature:**
+MiniMax-H3 supports powerful inline performance tags directly within `<d>[Language] ...</d>` blocks to direct vocal cadence, micro-pauses, emotional tone, and word emphasis far more effectively than descriptive text prose:
+
+| Tag | Function / Effect |
+| :--- | :--- |
+| `<pause>` | Short natural conversational pause (~0.3s - 0.5s) |
+| `<long pause>` | Extended thoughtful or dramatic silence (~1.0s - 1.5s) |
+| `<breath>` / `<inhale>` / `<exhale>` | Audible natural inhalation or exhalation before/during speech |
+| `<deep breath>` / `<catches breath>` | Dramatic emotional breath intake |
+| `<sighs>` | Audible vocal sigh before delivering line |
+| `<whisper>` | Low-volume, breathy whisper delivery |
+| `<softer>` | Subtle reduction in vocal intensity/volume |
+| `<stutter>` | Hesitant, stammering vocal delivery |
+| `<laughs>` / `<chuckle>` | Injected chuckling/laughing while speaking |
+| `<i>word</i>` | **Vocal emphasis on a specific word or phrase** (e.g. `<d>[English] I did <i>not</i> authorize that protocol.</d>`) |
+
+### Rule 55: Editorial Assembly & Montage as the True Engine of Film Consistency (September 2026)
+**The Core Philosophy:**
+Treat MiniMax-H3 not as a one-click scene generator, but as a **tool for producing individual cinematic shots**.
+1. **Short, Controlled Generations**: Build movies from a high volume of relatively short, focused generations (5–12s) rather than attempting to force 30–60 second scenes in a single take.
+2. **Simplified Prompts**: Prompts work best with **one clear action, clear character position, clean reference frame, clear camera command, and focused dialogue**. Change one variable at a time when troubleshooting.
+3. **Editing Hides Inconsistencies**: Cut between complementary angles (Shot -> Reaction -> Reverse -> Medium). Carrying dialogue across cuts (J/L cuts) and cutting before a generation starts to drift or degrade makes the overall film feel seamless and professionally directed.
+
 ---
 
 ## 3. Multi-Scene Movie Production Workflow (`docs/movies/movieXYZ/`)
@@ -772,6 +894,15 @@ N/A
 - [ ] Are complex sequence moments decomposed into focused micro-clips (alert -> silent cascade -> status -> deduction) rather than crammed into a single rushed take (Rule 44)?
 - [ ] For a new custom-LoRA character: 15× 5s R2V lock-in with rotating 8 refs and mix-and-match 0.40–0.50 on each LoRA, then a same-location C2V continuity chain (Rule 45 / CHARACTER_LOCKIN_DRILL.md)?
 - [ ] Is 2-LoRA + 8-ref strength the character's confirmed preset (Natalia 0.45/0.45), not a blind 0.80 default (Rule 34)?
+- [ ] Has dialogue duration been calibrated at low resolution (0.4 MP) before committing to 2 MP generation, ensuring speech does not race or drag (Rule 47)?
+- [ ] Are hero dialogue close-ups committed at ~2 Megapixels / Full HD to eliminate the plastic "AI video" look (Rule 48)?
+- [ ] Is the starting reference image high quality with correct composition, lighting, and wardrobe, acting as the visual authority (Rule 49)?
+- [ ] For static shots, are all five camera parameters (distance, scale, framing, background, position) explicitly locked to prevent slow drift (Rule 50)?
+- [ ] In single-character close-ups/reactions, are off-screen characters strictly omitted from descriptive prose to prevent blending or duplicate spawning (Rule 51)?
+- [ ] Is off-screen dialogue separated into clean dialogue capture + silent reaction shot, combined via J-cut / L-cut in editing (Rule 52)?
+- [ ] Do silent reaction shots carry the 4-point anatomical silence protocol (resting mouth posture, lips stationary, jaw still, no mouthing) (Rule 53)?
+- [ ] Are inline performance tags (`<pause>`, `<breath>`, `<whisper>`, `<i>word</i>`) used inside `<d>` for vocal nuances and word emphasis (Rule 54)?
+- [ ] Is the production broken into short, focused shots (5–12s) with simple prompts and dynamic edit pacing rather than monolithic takes (Rule 55)?
 
 ---
 

h3-center/docs/examples/test_billieeilish_turbo8_3310245_singing_birds_of_a_feather_8s_00001_.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98d19fe0490dd21a34a8776a5d8c363625a6528259aa8d902063880acbfb1abe
+size 2539815

h3-center/docs/examples/test_billieeilish_turbo8_3310245_speech_voice_ref_8s_00001_.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d6e85e9b70e97defe6fb4667434247c1faf5e78bcba661355c01fc3dce67615
+size 1752182

h3-center/docs/examples/test_dual_billie_miley_turbo8_3310245_simultaneous_twoshot_6s_00001_.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a1aeef79fe1a8ed6874a6ce0855eccaa912e9a75cad36a06d7e0f0712e12e42c
+size 849825

h3-center/docs/examples/test_dual_billie_miley_turbo8_3310245_singing_duet_madonna_6s_00001_.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00d4cb019a9f3e8b40d6229cbcc5f276befb03798c96803530364cfeb91e9b0d
+size 1236130

h3-center/docs/examples/test_mileycyrus_turbo8_3310245_singing_jolene_8s_00001_.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f26c5e3466344ac34e6001c0c74b0117dc47bcdd3df29e64ca43e02a7b43d4e
+size 2411118

h3-center/docs/examples/test_mileycyrus_turbo8_3310245_speech_voice_ref_8s_00001_.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7db833e8171d8a234c747726d075327811dd42bdba2e41b40f716f36220c1a3
+size 1762374