lucidml's 800M causal diffusion model turns images into playable characters at 60+ fps
Local consumer GPU model creates controllable characters from any image with improved stability.
This is a sequel to the previous 500M model, now 800M parameters. It runs comfortably on consumer GPUs. Context is increased to 12 latent frames. The weird flashes are gone, stability is much better, though consistency is horrible. The architecture is the same—just fattened the MLP. The denoiser is trained from scratch with diffusion forcing. It uses causal diffusion: a denoising loop for each frame added to a KV cache, with a sliding window to stay within trained context (20-30 latent frames). The 500M model gets over 60 fps on an RTX 5090. Posted to /r/lucidmlx.
- 800M parameter causal diffusion model runs on consumer GPUs; RTX 5090 achieves 60+ fps on the 500M variant.
- Uses diffusion forcing with a KV cache and sliding window to maintain temporal consistency across 12 latent frames.
- Trained from scratch; stability improved over the previous version but character consistency across frames remains a known issue.
Why It Matters
Democratizes real-time, local character generation from images, enabling indie creators to build interactive avatars without cloud costs.