AnyGameRenderer LoRA raises entity match on MiniMax-H3
October 21: Hugging Face reports measured gains—temporal match from 50.9% to 62.9% and missed count changes from 34.7% to 21.3% on 59 cases with 753 entities.
A LoRA on 33B MiniMax-H3 raises temporal entity match from 50.9% to 62.9% on 59 held-out game cases.
AnyGameRenderer is a video-to-video generative renderer that turns gray blockouts and other simple game footage into detailed visuals while keeping layout, camera, and each entity’s position and timing. It takes plain video, not engine buffers or class labels, and is a LoRA fine-tune of MiniMax-H3 ref2va, a 33-billion-parameter reference-video model. Training uses sim10k paired clips from 38 GEAR agent-made games, Unreal Engine crowds, and real gameplay. On 59 held-out cases with 753 entities, the authors report temporal match rising from 50.9% to 62.9% and missed count changes falling from 34.7% to 21.3% versus the base model.
- LoRA fine-tune of MiniMax-H3 ref2va, a 33B reference-video model.
- Training uses sim10k pairs from 38 GEAR games, Unreal, and gameplay.
- Inputs include blockouts, primitives, semantic maps, and wireframes.
- Entity benchmark has 59 held-out cases and 753 entities scored against boxes.
- Temporal match rises from 50.9% to 62.9% versus the base model.
- Missed count changes fall from 34.7% to 21.3%.
What each outlet adds
beyond the shared facts- Hugging Face daily papersEntity scores, 753-entity benchmark, and sim10k training mix.
- Products
- GEARUnreal Engine
Coverage
oldest first · one row per article · UTC- · 6d ago
Hugging Face daily papersFirst report
SolarRenderer adapts MiniMax-H3 for four-step structure-conditioned video
Two LoRAs keep a frozen 33B backbone and render about 12.3 times faster than 50-step sampling.
- · 4d ago
AnyGameRenderer restyles game footage while keeping entities
A LoRA on 33B MiniMax-H3 raises temporal entity match from 50.9% to 62.9% on 59 held-out cases.