VLM-all-inclusive (OmniVLM)
an omni-model built from scratch that sees, hears, speaks, and lip-syncs
3 min readupdated
A research project from hkblockchaintechnology: a unified AI model trained completely from scratch on a single GPU without off-the-shelf pretrained backbones (like ready-made ViT or HiFi-GAN). The model takes an image and a question, generates a text answer, voices it, and renders a talking-head video with synchronized lip-sync, while also supporting speech recognition on the input. All multimodal capabilities are built around a shared vocabulary (8,011 words + 39 ARPAbet phonemes) and a single weight trunk, where speech heads are trained in multi-task mode alongside vision. Best suited for multimodal architecture researchers and enthusiasts interested in full end-to-end training pipelines for omni-models.
Key features
- Full pipeline from scratch: vision (ViT, 10 layers, 128px scenes), language core (16-layer transformer with RoPE and cross-attention), hearing (CTC phoneme recognizer), voice (FastSpeech-style decoder), and vocoder (Vocos-style with ConvNeXt and ISTFT) are all written and trained without external pretrained modules.
- Unified token space: shared weights process vision-language tasks while simultaneously driving speech synthesis and articulation, without splitting the model into isolated copies.
- Built-in lip-sync renderer: the model maps its own ARPAbet phoneme timeline to visemes and mouth, jaw, and eyebrow motion parameters, assembling the final
mp4video viaffmpeg. - On-GPU scene generation: the vision-language dataset (shapes, colors, sizes, counting, and spatial relationships on a 3×3 grid) is rendered by a vectorized rasterizer directly on the GPU to avoid CPU bottlenecks.
What you can do
- Explore multi-task omni-model architecture: inspect the training code for the unified weight trunk, where the CTC hearing module self-labels data for the speech decoder without an external forced aligner.
- Swap synthetic scenes for real data: following the author's instructions, you can replace the
omni/scenes.pymodule with a real dataset for VQA or image captioning and fine-tune the model. - Test end-to-end video response generation: run the
OmniSystem.test_to_movie()method to generate a narrated video answering a question about a scene in ~2.5 seconds on a single GPU.
Example
python -m omni.train --stage vlm --preset xl --steps 8000 --batch 384
python -m omni.train --stage ear --steps 4000 --batch 24
python -m omni.train --stage align
python -m omni.train --stage speech --steps 6000 --batch 24
python -m omni.train --stage vocoder --steps 4000 --batch 16
Read more
DetailsTrained on an RTX PRO 6000 Blackwell (100 GB VRAM); the current snapshot has completed 251 steps for VLM (loss 0.1577), 503 steps for ear, and 2,251 steps for vocoder (the align and speech stages have not been run yet).
The speech module was trained on a single speaker from the LJSpeech-1.1 dataset (24 hours, CMUdict).
Limitations: the visual corpus currently consists only of synthetic geometric shapes, lip-sync uses an analytical 2D renderer instead of photorealistic generation, and there is no RLHF or moderation.










Comments