Anime / Stylized Generation

1. Install ComfyUI via Stability Matrix.
2. ComfyUI Manager → Install Models → search and download:
   • "noobai-xl-vpred-v10" (~7 GB — the leading open-weight anime model on SDXL architecture)
   • Or "animagine-xl-4.0" (~7 GB — strong alternative with good prompt adherence)
3. Load the default SDXL workflow. Set the model to NoobAI XL.
4. Use anime-specific prompting syntax: "1girl, solo, long hair, blue eyes, school uniform, standing, cherry blossoms, masterpiece, best quality, detailed."
5. Resolution: 1024×1024 or 832×1216 (portrait). Steps=25-30, CFG=5-7, sampler=DPM++ 2M Karras.
6. First image in 8-15 seconds on 12 GB GPU. Add LoRAs (from CivitAI, download .safetensors → place in ComfyUI/models/loras) for character consistency and specific styles.

Used RTX 3060 12 GB (~$200-250, see /hardware/rtx-3060-12gb). Runs NoobAI XL or Animagine XL at 1024×1024 in 8-15 seconds — fast enough for iterative prompting ("regenerate with different hair color"). SDXL-based anime models are the sweet spot: 12 GB runs them comfortably. Pair with Ryzen 5 5600 + 32 GB DDR4 + 1TB NVMe (anime LoRAs and models add up fast, you want storage). Total: ~$390-440. For SD 1.5-based anime models (AOM3, Anything V5): they run on 4 GB cards at 3-6 seconds per image but with lower detail quality.

Used RTX 3090 24 GB ($700-900, see /hardware/rtx-3090). Runs NoobAI XL at 3-5 seconds per image, can batch 4 images simultaneously (12-15 seconds for 4). Also runs Flux + anime LoRAs at 8-15 seconds for higher-quality outputs. For consistent character generation: the extra VRAM allows loading multiple LoRAs (character + outfit + pose) simultaneously without offloading. Pair with Ryzen 7 7700X + 64 GB DDR5 + 2TB NVMe. Total: $1,800-2,200. RTX 4090 ($2,000) at 1-2 seconds per anime image — fastest iteration.

The mistake: Using a realistic model (Flux Dev, SDXL base) and adding "anime style" to the prompt, expecting it to produce anime. Why it fails: Realistic models are trained on photographs — they produce photorealistic images with anime-like coloring, not actual anime. The anatomy, proportions, line art, and cel-shading won't match anime aesthetics. The fix: Use purpose-built anime models: NoobAI XL, Animagine XL, Pony Diffusion V6 XL. These are fine-tuned on Danbooru and other anime image boards. They understand anime-specific tags (1girl, solo, blue archive, bangs, thighhighs) and produce authentic anime-style art with correct proportions, line art, and cell shading. Anime is a distinct visual domain — use domain-specific models.