IP-Adapter-FaceID variants: face-ID embeddings, settings, a Portrait snippet bug and research-only use

IP-Adapter-FaceID (experimental) replaces the CLIP image embedding of IP-Adapter with a face-ID embedding from a face recognition model, and uses a LoRA to improve identity consistency, so text prompts can place a given face in many styles. Variants: FaceID-Plus adds a CLIP image embedding of the face for structure; FaceID-PlusV2 makes that structure weight adjustable (shortcut=True with s_scale); FaceID-SDXL and FaceID-PlusV2-SDXL are experimental SDXL versions; FaceID-Portrait uses no LoRA or ControlNet and takes several face images (5 by default) for higher similarity. Face embedding: insightface FaceAnalysis(name="buffalo_l", providers=["CUDAExecutionProvider", "CPUExecutionProvider"]), app.prepare(ctx_id=0, det_size=(640, 640)), then faces[0].normed_embedding as a 1 x 512 tensor. For Plus, also crop an aligned face with face_align.norm_crop(image, landmark=faces[0].kps, image_size=224). Generation (code from github.com/tencent-ailab/IP-Adapter, ip_adapter package): SD1.5 examples use SG161222/Realistic_Vision_V4.0_noVAE with the stabilityai/sd-vae-ft-mse VAE and DDIMScheduler (1000 train steps, beta 0.00085 to 0.012, scaled_linear, clip_sample=False, set_alpha_to_one=False, steps_offset=1), then IPAdapterFaceID(pipe, "ip-adapter-faceid_sd15.bin", device).generate(prompt, negative_prompt, faceid_embeds, num_samples=4, width=512, height=768, num_inference_steps=30, seed). A separate loader instead loads and fuses ip-adapter-faceid_sd15_lora.safetensors explicitly. Plus uses IPAdapterFaceIDPlus with image encoder laion/CLIP-ViT-H-14-laion2B-s32B-b79K and ip-adapter-faceid-plus_sd15.bin or -plusv2_sd15.bin. SDXL uses IPAdapterFaceIDXL with SG161222/RealVisXL_V3.0, ip-adapter-faceid_sdxl.bin, 1024x1024, 30 steps, guidance_scale=7.5. Portrait uses ip-adapter-faceid-portrait_sd15.bin with num_tokens=16, n_cond=5 at 512x512. The examples pass safety_checker=None. Correction: the card's Portrait snippet loops over five file names but reads "person.jpg" each time, so all embeddings come from one image, and its torch.cat line is mis-indented. Read each listed file instead. Limitations: imperfect photorealism and identity consistency; generalization limited by training data, base model and face recognition model. Licence condition: because InsightFace pretrained models are available for non-commercial research only, IP-Adapter-FaceID models are released exclusively for research and not for commercial use. Based on the h94/IP-Adapter-FaceID card pinned to revision 43907e6f44d079bf1a9102d9a6e56aef7a219bae and observed on 2026-10-27. Nothing was run here; the IP-Adapter repository code and the InsightFace licence text were not reviewed, and PlusV2-SDXL usage is not shown in the card. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/h94/IP-Adapter-FaceID/raw/43907e6f44d079bf1a9102d9a6e56aef7a219bae/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 394:633, 698:797, 1455:1655, 1951:2091, 9266:9500, 10367:10441, 11015:11237; exact excerpt SHA256 fb9fbc29d9cf9b621d1bbd9e6ee3476a7b082d2da26158e7979cec4df4ef224f,7c545cb57cbc9fba5ca7da49cedb88d5844c0307c24ef33e7ab9ed3e4ddfc4f7,a2714dcaa4612a8abc21fffb8b014d1d10b613202f6f7cf9b8"
    }
  ]
}