凤希AI - 多视图生成器

FxAiQwenEditMultiView turns a single headshot of a person into a four-panel character sheet: a close-up on the left, and three full-body views (front, left profile, back) stacked on the right. It's the pack's take on the 2026 character-sheet workflow - reference image in, plain-language instruction, no ControlNet - built directly on Qwen-Image-Edit. The 四视图V2 example workflow in the repo is exactly this node wired end to end.

The clever part is the pre-processing. Before anything touches the model, the node runs OpenCV's Haar face detector on your 人物头像 (avatar IMAGE) input, finds the largest face, crops to it with a margin, and pastes it onto a pure white background. That background stripping is deliberate: it stops the edit model from copying the reference photo's background, clothes, or hairstyle into the sheet - only the facial features are meant to survive. It then builds a long English prompt that hard-codes the canvas layout (25% close-up, 75% three stacked full-body views), locks the facial identity to the reference, and makes everything else (hair, outfit, body, accessories) follow your 用户提示词 instead. A matching English system prompt enforces the same rules, and both positive and negative conditioning get the reference latents attached so the model holds the face while everything else is free to change.

Inputs: clip, vae, width/height (default 960×1280, aligned up to a multiple of 8 internally), 用户提示词 (STRING, what the character should look like), and 人物头像 (IMAGE). Outputs: positive, negative (CONDITIONING) and latent (LATENT) - straight into a KSampler.

Two setup notes. First, this node requires cv2 (opencv-python) for the face detection - it imports it at module load, so if that Python package is missing, this node won't even register. pip install opencv-python in your ComfyUI environment if the node doesn't appear. Second, it needs the Qwen-Edit model stack, which the pack does not download: qwen_image_edit_2511_bf16.safetensors (or fp8/GGUF) via UNETLoader type qwen_image, qwen_2.5_vl_7b_fp8_scaled.safetensors via CLIPLoader type qwen_image, and qwen_image_vae.safetensors. At 20B, quants are the realistic path on consumer cards.

The honest limit is the model's: Qwen-Edit re-emits the whole frame, so faces drift and layouts can wobble - the 2511 revision improved but didn't cure it, and the 2026 community answer is "consistency lora, then inpaint if needed" (character-consistency.md). Also be aware the Haar detector is old-school: good frontal faces work great, heavily angled or tiny faces may not be found, in which case it falls back to the whole image on white - check your output.

It's under 凤希AI/图片, part of fxai-toolkit (凤希AI / fxai.site):

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

or ComfyUI Manager → search "fxai-toolkit", restart, and pull the three model files plus opencv. One good headshot and it does the rest.