Skip to main content
Page is under construction.

Check the github issues for ways to contribute! Or provide your feedback in this quick form
This page answers one question:
What models can run on Livepeer, and why might some not?
It lists all known model families commonly used via ComfyUI, with compatibility ratings for Livepeer’s real-time, GPU-worker constraints. Nothing here implies that a listed model is officially supported or pre-loaded on the network — it reflects whether a model’s execution shape fits Livepeer well.

Legend

  • ✓ Likely runnable — fits real-time / GPU-worker constraints
  • ⚠ Conditional — depends on latency, VRAM, orchestration, or batching
  • ✗ Not suitable — design mismatch: stateful, CPU-bound, or non-deterministic

1. Diffusion Models (Image / Video)

Stable Diffusion family

ModelFitNotes
Why blocked (DeepFloyd): VRAM pressure, multi-stage graphs, inference latency.

Video diffusion models

ModelFitNotes
Why blocked (batch video): temporal state, batch-only execution, non-real-time.

2. Control & Conditioning Models

ControlNet

ModelFitNotes

T2I / I2I Adapters

ModelFitNotes

3. Encoders, VAEs, and Latents

ModelFitNotes

4. Vision Models (Non-Diffusion)

Detection / Segmentation

ModelFitNotes

Depth / Geometry

ModelFitNotes

5. Face, Pose & Human Models

ModelFitNotes

6. Audio & Music Models

ModelFitNotes
Why blocked: long context windows, non-frame-based execution.
For real-time audio workloads (live ASR, live translation, streaming transcription), see Workload Fit → ASR pipeline examples. These use Whisper or similar and are excellent fits.

7. Multimodal & VLMs

ModelFitNotes

8. LLMs (Text-Centric)

ModelFitNotes
Why blocked: token streaming, memory residency, orchestration mismatch.

9. 3D / NeRF / World Models

ModelFitNotes

10. Utility / Pre/Post Models

ModelFitNotes

Core takeaway

ComfyUI can orchestrate almost any PyTorch model. But:
  • Livepeer favours stateless, frame-based, deterministic inference
  • Long-running, stateful, or batch-only models are fundamentally incompatible
  • Real-time video imposes hard physics limits, not software ones
This matrix is intentionally conservative. If your model doesn’t appear here, apply the Workload Fit decision tree to evaluate it.

See also

Last modified on March 2, 2026