凤希AI - 三视图生成器

FxAiQwenEditThreeView is the simpler sibling of FxAiQwenEditMultiView: same family of job - generate a standard three-view character reference (front, left, back, full body) on a pure white background using Qwen-Image-Edit - but with the face-detection preprocessing removed. No OpenCV dependency, no background-stripping pass; you hand it an optional reference 人物头像 and a 用户提示词, and it builds the conditioning. If the MultiView node is the "make it look like this specific person" version, this one is the "make a clean three-view sheet" version.

Mechanically it's the standard Qwen-Edit encode recipe. If you supply a 人物头像, each image becomes a numbered <|vision_start|><|image_pad|><|vision_end|> block in the prompt, gets rescaled to a ~384-pixel baseline for the vision encoder, and gets VAE-encoded at a ~1024 baseline into reference latents attached to the positive conditioning - the appearance-control half of Qwen-Edit's dual encoding (see the qwen-image-edit.json panel). The prompt then instructs the model to preserve the person's facial identity and build three full-body views in a fixed order (front → left → back), full body in frame, no cropping, no extra views, pure white background. The system prompt does the same frame, and the negative is the standard Chinese anti-artifact block (ugly, blurry, extra limbs, bad hands, text, watermarks…).

Inputs: clip, vae, width/height (960×1280 default, aligned to multiples of 8), 用户提示词 (required STRING), and 人物头像 (optional IMAGE - leave it unplugged and it still works, generating a sheet from text alone). Outputs: positive, negative (CONDITIONING), latent (LATENT) - a blank tensor at the target resolution, ready for KSampler. The repo's 三视图V4 workflow is this node wired with the standard Qwen-Edit model stack plus a stack of community LoRAs (a Lightning 8-step LoRA, a "Cinematic" LoRA, realism LoRAs).

The model requirement is the same as every Qwen-Edit node in this pack, and it's not bundled: qwen_image_edit_2511_bf16.safetensors (or fp8/GGUF - UNETLoader type qwen_image), qwen_2.5_vl_7b_fp8_scaled.safetensors (CLIPLoader type qwen_image), and qwen_image_vae.safetensors. Search those names on HuggingFace. bf16 wants a big card; quants are the practical route.

The tradeoff versus the MultiView node is worth stating plainly: without the face-crop preprocessing, whatever's in the reference image - background, clothes, hair - is fair game for the model to copy, so feed it a clean headshot if you want identity without costume bleed. And as with the whole Qwen-Edit family, faces can drift and the model re-emits the whole frame; the 2026 playbook is a consistency LoRA plus an inpaint pass if the face wanders (character-consistency.md). It's in 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. If you just want a clean character reference sheet without the extra face plumbing, start here.