凤希AI - 人物面部特征图

You don't get to write a prompt here. That's the point, and it's the first thing to know about FxAiQwenFaceFeature (凤希AI - 人物面部特征图). Feed it a photo of a person - plus a Qwen CLIP and VAE - and it hands back conditioning and an empty latent that make Qwen-Image-Edit render a frontal head close-up of that person on a pure white background.

Why you'd want that: a clean, white-background face with the identity locked is the asset every consistency workflow downstream eats - a LoRA dataset starter, a character-sheet panel, the reference your next edit gets handed. The 2026 answer to "how do I keep the same character" is "edit the one good image you already have", and every one of those pipelines begins here.

What it's actually doing

Under the hood it's ComfyUI's own Qwen-Image-Edit text-encode step with the prompt baked in. It resizes each connected image to an area of 宽度 × 高度 (aspect preserved, sides snapped to multiples of 32), feeds them to the Qwen2.5-VL vision tower as <image1>/<image2> tokens, VAE-encodes the same images and attaches them as reference_latents on both conditioning tensors, then zeroes out a latent at exactly 宽度 × 高度. Those are the three things it returns.

Which prompt you get depends on one connection. With 参考头像 wired up, the instruction is "keep <image1>'s facial features, put <image2>'s clothes and hairstyle on it, frontal head close-up, pure white background". Without it, "generate this person's frontal head close-up feature image, consistent with <image1>, pure white background" - and 人物图片 becomes <image1> instead of <image2>. There's a hardcoded Chinese negative junk-token list, and no field to change any of it.

Inputs and outputs that matter

Required: clip (a Qwen-Image text encoder - the pack's own workflow loads qwen3vl_8b_int8 with type qwen_image), vae, and 宽度/高度, both defaulting to 1024.

Optional, and where you'll actually make mistakes: 参考头像 is the character reference that becomes <image1>, and 人物图片 supplies the clothes and hair as <image2>. Swap them and the prompt does the reverse of what you wanted. Connect only 人物图片 and you get the no-avatar path - a straight face sheet from that one photo.

宽度/高度 does two jobs: it's the output resolution and the resize budget for both reference images. The output is always exactly 宽度 × 高度 whatever shape your input was, so square 1024×1024 is fine for a sheet and your references don't need to be square.

On the way out, 正向条件 and 负向条件 are CONDITIONING for the KSampler's positive and negative, and 潜空间 is a LATENT that goes straight into its latent input - no Empty Latent node needed, this one already made it.

Installing

ComfyUI Manager: search fxai-toolkit, or 凤希 / FxAi if the registry is being shy. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/fxai666/fxai-toolkit
# restart ComfyUI

The pack ships no model files - it's nodes only. You bring Qwen-Image-Edit (base, 2509 or 2511), its Qwen2.5-VL text encoder, and the Qwen VAE. There's no requirements.txt, but __init__.py pip-installs soundfile and psutil on import (for the audio nodes, not this one), so the first launch wants internet. If it silently gives up, install them into ComfyUI's Python yourself:

python -m pip install soundfile psutil

Where people get burned

Nothing comes out that looks like your person. Both image inputs are optional, so the node runs happily with neither - no reference latents, just a fixed prompt and a random image. And reference_latents only means something to the Edit checkpoints; plug the text-to-image Qwen-Image model in and your references are ignored.

Wrong VAE. Wrong channel count, wrong VAE: noise or flat colour, not a subtly wrong image. And if the close-up comes back plasticky - the Qwen-Image VAE has a reputation for over-smoothing, Qwen froze that encoder and only fine-tuned the decoder, so the Wan 2.1 VAE shares the latent space and drops straight in.

The negative prompt is probably doing nothing. The pack's own 四视图 workflow samples at CFG 1.0, and at CFG 1 there's no unconditional pass, so a negative conditioning is inert - not weak, unused.

The likeness isn't quite them. Don't blame your settings first. Identity drift on faces is the documented weak spot of the whole Qwen-Image-Edit line, worst on faces the model hasn't seen in training. 2511 claims improved character consistency; if you can't get there, the honest fix is a character LoRA or a masked second pass on the face, not more steps.

You want to change something. You can't. Background, framing and outfit are fixed strings. The same pack's FxAiQwenImage21Edit takes real prompt inputs when you need that control; this node is a preset, and a good one if you treat it as one.

Housekeeping: fxai-toolkit registers around a hundred Chinese-labelled nodes, and it ships an HTTP self-update endpoint that runs git fetch + reset --hard in its own folder - don't press that if you keep local edits.