01
CheckingSpeech + voice cloning
Built-in speech works on CPU. Voice-sample cloning also requires acceptance of Kyutai's gated model terms and an HF token.
IDEAVERSE · ISOLATED MODEL LAB
One surface. Eleven model proofs.
Every output below is generated on the dedicated deployment VM—not by the main Ideaverse application.
Checking hardware…
02
CheckingText to image
Sana Sprint 0.6B creates a 1024px image with its required two-step SCM pipeline.
03
CheckingText to video
CPU generation is live at 832×480, nine frames and three steps. Expect roughly 14 minutes on this four-vCPU VM; a GPU enables the full 81-frame profile.
Play verified CPU sample →04
CheckingVideo to audio
MMAudio generates synchronized sound effects. The VM uses its small CPU checkpoint; a GPU switches to large 44k v2.
05
CheckingDocument understanding
The official Baidu model and 6.4 GB checkpoint are staged; its official runtime requires NVIDIA CUDA.
06
CheckingFast OCR
PP-OCRv6 Small detects text regions and recognizes their content efficiently on CPU.
07
CheckingRealtime talking avatar
SoulX, Wav2Vec2, LoRA, audio projection and Lite TAE weights are staged; inference requires NVIDIA CUDA.
08
CheckingUnified audio scenes
MiDashengLM-Gen creates speech, music, sound effects and ambience together. Its 11.54 GB official checkpoint runs locally on this CPU VM.
Play verified CPU sample →09
CheckingIndexTTS 2.5 voice cloning
The latest IndexTTS clones one reference voice, supports five languages and exposes speaking-duration control. CPU synthesis is slow but fully functional.
Play verified CPU sample →10
CheckingMiniMax Music 3
The official runtime requires two CUDA GPUs. This lab instead builds the experimental audio.cpp CPU port and converts the official model to memory-saving Q4_K GGUF components.
Play verified CPU sample →11
CheckingAuK-Flash unified speech
AuK-Flash follows natural-language instructions for speech generation, voice-aware editing, enhancement and separation. The official distilled model uses four inference steps; this lab runs a serialized CPU proof, so generation may be slow.