TTV Studio
Online

Generate Video

Transform conceptual text descriptions into high-fidelity synchronized cinematic videos.

Suggestions:
0 chars • 0 words
▼
15 seconds
5s 15s 30s 45s 60s
▼
Advanced Generation Settings
▶
▼
▼

Video Preview

Ready

No video generated yet

Enter your text prompt on the left and click "Generate Video" to render your cinema-grade video.

My Videos & History

Browse, replay, and manage previously rendered text-to-video deliverables.

AI Video Models

Inspect configured generative backbones, LoRA adapters, and multimodal inference pipelines.

All Checkpoints Loaded
Active Backbone

SpatialTemporal TTV Model

Multi-scale spatial-temporal latent diffusion model with 3D convolutions and Cross-Attention story conditioning.

Type: Diffusion Backbone
Latent Channels: 64-dim
Temporal Attention: Enabled (8-60 frames)
Resolution: 720p / 1080p HD
Loaded in Memory
LoRA Adapter

TTV LoRA v001 (Cinematic)

Low-Rank Adaptation trained on cinematic dataset pairs. Enhances photographic texture, depth, and smooth camera pan motion.

LoRA Rank (r): 8
LoRA Alpha: 16
Base Model: SpatialTemporal TTV
Target Modules: temporal_conv, attn
Checkpoint Available
Subsystem Providers

Multimodal Pipelines

Integrated prompt expansion, keyframe synthesis, and text-to-speech audio orchestration.

Story Director: Local LLM / Story Service
Keyframes: Vision Consistency Service
Voice Engine: TTS Speech Synthesizer
Video Assembler: FFmpeg High-Performance
Production Ready

Model Training & Fine-Tuning

Configure dataset manifests, adjust LoRA hyperparameters, inspect checkpoints, and launch training jobs.

Trainer Ready

Dataset Management

Click to select or drag and drop dataset manifest

Supports JSON manifest, CSV, or video directory archive
45 Video Samples
1280x720 Resolution
80% / 20% Train / Val Split
manifest.json Format

Hyperparameters & Method

Training Execution Worker

Worker Ready
[System] Training environment initialized with PyTorch.
[Dataset] Found 45 video deliverables in generated/videos.
[Model] SpatialTemporalTTVModel loaded. LoRA rank=8 configured.
[CLI] Execute: python training/run_training.py --config training/configs/smoke_test.yaml

Evaluation Benchmarks

Training Loss (MSE) 0.038 Target: < 0.050
Validation Loss 0.044 Target: < 0.060
Fréchet Video Distance (FVD) 142.6 Lower is better
Temporal Consistency (SSIM) 0.892 Target: > 0.850

Registered Checkpoints

ttv_lora_epoch_5.pt Epoch 5 • 14.8 MB • LoRA Rank 8
best_model.pt Best Validation Loss (0.041) • 14.8 MB

RL Preference Feedback Studio

Human-in-the-loop Direct Preference Optimization (DPO) and Composite Reward scoring for continuous model refinement.

Iteration 1 / 2
Active Evaluation Prompt: "A small robot explores a futuristic city at sunset, cinematic lighting."
Candidate A (Base Engine) Score: 0.82
Aesthetics: 0.84
Motion: 0.80
Alignment: 0.82
Candidate B (LoRA Adapted) Score: 0.91
Aesthetics: 0.92
Motion: 0.89
Alignment: 0.91

Composite Reward Weights

Weights defined in training/configs/rl_loop.yaml

Aesthetic Quality: 0.40 Motion Smoothness: 0.30 Prompt Alignment: 0.30

Recent Feedback Submissions

Studio Settings & System Readiness

Inspect underlying microservice connectivity, hardware acceleration, and provider configuration.

Microservice Provider Readiness

LLM Orchestration: local
Visual Keyframe Synthesizer: local
Temporal Video Motion Engine: opencv
Voice & Audio Synthesis (TTS): local
FFmpeg Media Encoder: Checking...

Governance & Access Token

The Unified Engine enforces governance token verification for secured generation requests.