凤希AI - 千问角色融合控制器

FxAiQwenEditEnhanced is the "角色融合" (character fusion) encode node: it's how this pack drives Qwen-Image-Edit for multi-person and character-control jobs. If the pack's 万图融合 or character workflows need to take a reference of a person, hand it a list of extra images, and edit everything into one scene, this is the node that builds the conditioning. It's not a sampler - it's the front half that produces the positive, negative, and latent you then run through KSampler.

The mechanism is Qwen-Edit's dual-encoding done properly. It takes a clip (a Qwen-Image-VL CLIP loader) and your 用户提示词/负面提示词, then assembles an image-part prompt: every image in 人物列表 (people, labeled 人物1, 人物2…) and 图片列表 (general images, labeled 图1, 图2…) gets a <|vision_start|><|image_pad|><|vision_end|> token block so the CLIP sees them as vision input. Each image is also rescaled and VAE-encoded into a reference latent at a ~1024-pixel baseline, and those reference latents are attached to the positive conditioning - that's the "appearance control" half of Qwen-Edit's design (see the qwen-image-edit.json panel). The text goes through Qwen's chat template with a built-in system prompt that tells the model to describe every input image then strictly obey your edit instruction. The 负面提示词 has a generous Chinese fallback (ugly, blurry, bad anatomy, extra limbs…) if you leave it empty, so the node never emits an empty negative.

Inputs: clip (required), 用户提示词 and 负面提示词 (both STRING), width/height (960×1280 default, 512–4096 step 8), optional vae, 人物列表 (IMAGE), 图片列表 (IMAGE). Outputs: positive (CONDITIONING), negative (CONDITIONING), latent (LATENT - a blank latent at width/8 × height/8, ready for the sampler).

Before it does anything useful you need the model files, and this is where beginners stall. The pack does not download them. You need three things in your models/ folders: the Qwen-Image-Edit diffusion model (qwen_image_edit_2511_bf16.safetensors or an fp8/GGUF quant - UNETLoader with the qwen_image type), the Qwen2.5-VL 7B text encoder (qwen_2.5_vl_7b_fp8_scaled.safetensors, CLIPLoader type qwen_image), and qwen_image_vae.safetensors (VAELoader). The pack's example workflows list exactly those filenames, so search those on HuggingFace. At 20B, bf16 is a 24GB+ card proposition - the community runs fp8 and GGUF quants, which is the realistic path on consumer VRAM.

The honest warning from the KB: Qwen-Edit re-emits the whole frame, so pixels you didn't ask to change drift, and faces - especially faces the model hasn't seen - can change identity (character-consistency.md). This node handles the encoding side cleanly; it can't fix what the model does at decode time.

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 above. Then the 万图融合V4 example workflow will actually run.