凤希AI - 千问图片编辑

FxAiQwenEditEnhancedV2 is the general-purpose Qwen-Image-Edit encoder in this pack - the one you reach for when the job is "edit this image according to this sentence," not "fuse a specific character into a scene." It's the V2 cleanup of FxAiQwenEditEnhanced: it drops the separate 人物列表 input, makes vae required, adds an overridable 系统提示词, and does the same dual-encoding under the hood. If the V1 is the character-fusion controller, this is the everyday photo editor.

How it works is the Qwen-Edit encode recipe: every image in 图片列表 (IMAGE) gets a numbered <|vision_start|><|image_pad|><|vision_end|> block in the user prompt (labeled 图1, 图2…), each is rescaled to a ~384-pixel baseline for the vision CLIP and VAE-encoded at a ~1024 baseline into reference latents that ride along on the positive conditioning - the "appearance control" half of Qwen-Edit's dual encoding, per the qwen-image-edit.json panel. The whole thing is wrapped in Qwen's chat template (<|im_start|>), with a default system prompt that instructs the model to inventory every input image, strictly follow your edit instruction, discard anything you don't ask about, and fill the frame. 负面提示词 falls back to a long Chinese anti-artifact list if empty, so you get a solid negative without writing one. The latent output is a blank float16 tensor at height/8 × width/8.

Inputs: clip and vae (both required), width/height (default 960×1280, 512–4096 step 8), then optional 用户提示词, 负面提示词, 图片列表 (IMAGE), and 系统提示词 (STRING - set this to the 系统提示词 output of FxAiMultiViewGenerator to reuse its layout rules). Outputs: positive, negative (CONDITIONING) and latent (LATENT) - the complete feed for a KSampler.

Setup is the same as every Qwen-Edit node in this pack, and the pack does not ship the weights. You need three files in your models folders: the diffusion model (qwen_image_edit_2511_bf16.safetensors or a quant, UNETLoader type qwen_image), the Qwen2.5-VL 7B text encoder (qwen_2.5_vl_7b_fp8_scaled.safetensors, CLIPLoader type qwen_image), and qwen_image_vae.safetensors. Find those filenames on HuggingFace; the bf16 is heavy, so fp8/GGUF is the realistic route under 24GB.

The honest limits come from the model, not the node: Qwen-Edit re-emits the whole frame, so unedited regions drift, and real faces can change identity across a chain of edits - 2026 practice is "Qwen-edit with a consistency lora," then inpaint the face back if it wanders (character-consistency.md). Also, unlike V1, this node requires vae - feed it a missing VAE and it errors immediately.

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, then grab the three model files. Pair it with FxAiMultiViewGenerator for character-sheet prompts and you've got the pack's whole Qwen editing story in two nodes.