These are estimates. Real results vary ±10–20% based on your CPU, background processes, and optimization flags. Using --use-sage-attention in ComfyUI adds ~15–20% speed on RTX 30–50 series. Benchmarks sourced from Chimolog (WAI-NSFW v14, 832×1216, 28 steps, ComfyUI) and community testing.

Anima 2B key requirements: 8GB+ VRAM minimum (~7GB model). Recommended settings: 30–50 steps, CFG 4–5, ~1MP resolution. Speed estimates assume fp16 + SageAttention + TeaCache (ANIMA_BOOSTER). GTX and AMD/Intel cards cannot use ANIMA_BOOSTER and will be significantly slower than shown. Model by CircleStone Labs / Comfy Org. Non-commercial license.

▶ Lower upscaling burden: While per-image generation is slower than SDXL, Anima produces high-quality results at native resolution with far less need to upscale. In most workflows, upscaling passes can be skipped entirely — which saves significant time overall and narrows the real-world gap with SDXL.

Wan 2.2 key requirements: Minimum 64GB system RAM recommended (model offloads to RAM — 32GB causes SSD spillover and 2–3× slowdowns). 720p requires 16GB+ VRAM. Only GeForce GPUs run 720p reliably — AMD and Intel Arc fail with OOM errors at that resolution. Using Kijai/EasyWan22 workflow with Lightx2v LoRA (8 steps). Benchmarks from Chimolog Wan 2.2 benchmark.

MiniMax H3 FP8 requirements: The FP8 pruned model is ~21GB. Cards with 24GB+ VRAM run it fully in VRAM. 16GB cards offload ~5GB to system RAM (+~40% time). 12GB cards offload ~9GB (+~100% time) — workable but slow. Under 12GB is not practical. AMD and Intel Arc lack the tooling needed for FP8 + Turbo LoRA. Turbo LoRA v4 (4 steps) + SageAttention + EasyCache. RTX 5090 720p anchor: ~34s (aistudynow.com benchmark).