P-Image-OSS ComfyUI test reveals VRAM limits
Local ComfyUI benchmarks for Pruna AI's P-Image-OSS demonstrate 8-megapixel generation in 8 steps, with W4A8 quantized text-to-image peaking at 18.1 GiB VRAM. Complex multi-image editing workflows quickly exceed consumer GPU memory while inheriting MiniMax H3's territorial licensing restrictions.
Adapting MiniMax H3's video backbone for 8-step still image generation delivers impressive speed, but running it locally remains punishing on consumer hardware. Between steep memory ceilings for multi-image editing and strict geographic licensing barriers, practical adoption outside enterprise setups faces major headwinds.
- –W4A8 quantization enables 8 MP text-to-image inference within 18.1 GiB VRAM, narrowly fitting 24 GB consumer GPUs like the RTX 3090 and 4090.
- –Running 8-step diffusion without Classifier-Free Guidance (CFG) dramatically reduces step latency compared to traditional 30-step diffusion pipelines.
- –Multi-image editing with up to 8 reference frames rapidly exceeds 24 GiB VRAM, demanding aggressive layer offloading or dedicated cloud instances.
- –The underlying MiniMax H3 Community License explicitly excludes deployment in the US, EU, UK, and South Korea, significantly limiting commercial open-weights utility.
- –Heavy architectural overhead from the 32B text encoder and built-in Qwen prompt upsampler guarantees high fidelity at the cost of high base memory pressure.
DISCOVERED
1h ago
2026-11-22
PUBLISHED
2h ago
2026-11-22
RELEVANCE
AUTHOR
prompt48