想用家用等級顯卡在本地生成乾淨、穩定的 AI 插畫？這篇直接給你可匯入的 ComfyUI 工作流與參數建議，整個流程以 GGUF 量化版的 Qwen-Image 為核心，在 12GB VRAM 的環境中就能順跑 1024×768 單張生成。本文特點是流程精簡、全程不需 LoRA，並可直接使用中文提示詞，主要模型為 Qwen_Image-Q3_K_S.gguf，在效能與畫質之間取得良好平衡；文字編碼則採 Qwen2.5-VL-7B-Instruct-Q3_K_S.gguf，並搭配 qwen_image_vae.safetensors 做解碼輸出。注意！此模型是可以產生 NSFW 圖像的，請斟酌使用。

ComfyUI 的優勢在於以節點方式拼裝推論管線，整個流程「文字條件 → 取樣 → 解碼 → 輸出」都能一目了然，不僅方便除錯，也能快速更換模型或調整步數。以 Q3 量化模型為例，在 12GB 顯存上就能順暢運行，生成速度約一分多鐘即可完成，影像穩定度與邊緣表現也相當乾淨。若電腦顯存更充裕，可以嘗試提升步數，或改用更高等級的模型（如 Q4、Q5），畫面細節會更豐富。至於文字理解能力，若覺得模型對提示詞的掌握不足，也能升級文字編碼模型，不過這會額外佔用顯存並拉長生成時間。

環境配置與模型放置

請先完成 ComfyUI 的基本安裝並能開啟軟體介面（教學點我）。為了讀取 GGUF，我們僅需安裝 ComfyUI-GGUF 這組客製節點（提供 UnetLoaderGGUF 與 CLIPLoaderGGUF），安裝完成後務必重新啟動 ComfyUI，否則節點清單不會更新。

模型與路徑

Diffusion（GGUF）：Qwen_Image-Q3_K_S.gguf ［🔗 下載點］
📁 放置 → ComfyUI/models/diffusion_models/
Text Encoder（GGUF）：原檔 Qwen_Qwen2.5-VL-7B-Instruct-Q3_K_S.gguf，建議改名為
Qwen2.5-VL-7B-Instruct-Q3_K_S.gguf ［🔗 下載點］
📁 放置 → ComfyUI/models/text_encoders/
mmproj（GGUF）：Qwen2.5-VL-7B-Instruct-mmproj-BF16.gguf ［🔗 下載點］
📁 放置 → ComfyUI/models/text_encoders/
VAE：qwen_image_vae.safetensors ［🔗 下載點］
📁 放置 → ComfyUI/models/vae/

說明：為了在 12GB VRAM 上維持順暢，本文採用 Q3 量化版。若電腦顯存或 RAM 更充裕，可改用更高位寬或更大檔的模型，並把取樣步數拉高以換取更好的細節。

工作流重點與參數建議

整體流程非常直覺：以 UnetLoaderGGUF 載入 Qwen-Image（Q3），用 CLIPLoaderGGUF 載入 Qwen2.5-VL（Q3）；正、負向提示詞分別由兩個 CLIP Text Encode 節點處理，負向提示詞建議放入常見去污詞，如「畸形、模糊、低品質、偽影」等，影像尺寸由 EmptySD3LatentImage 產生潛空間圖，接著把條件、模型與潛變量送入 KSampler 進行取樣，最後透過 VAE Decode 與 SaveImage 輸出圖片。本文範例的基礎設定為 1024×768、步數 10、取樣器 res_multistep、排程 simple、CFG 在 1～4 之間微調；顯存吃緊時請維持單張批量。若想提高精緻度，可以把步數拉到 15～20，或在正向提示加入更具體的光影與構圖語彙。

📝 工作流（QwenImageGGUF.json）

將下列內容複製到文字編輯器，儲存為 QwenImageGGUF.json 後即可在 ComfyUI 匯入。

{"id":"a3b0f2e8-7d8f-4e1c-9b33-12d6f1c7a9cd","revision":1,"last_node_id":11,"last_link_id":11,"nodes":[{"id":2,"type":"CLIPLoaderGGUF","pos":[20,160],"size":[340,120],"flags":{},"order":0,"mode":0,"inputs":[],"outputs":[{"name":"CLIP","type":"CLIP","slot_index":0,"links":[1,2]}],"properties":{"cnr_id":"comfyui-gguf","ver":"1.1.4","Node name for S&R":"CLIPLoaderGGUF","models":[{"name":"Qwen2.5-VL-7B-Instruct-Q4_K_S.gguf","url":"file://text_encoders/Qwen2.5-VL-7B-Instruct-Q4_K_S.gguf","directory":"text_encoders"}]},"widgets_values":["Qwen2.5-VL-7B-Instruct-Q3_K_S.gguf","qwen_image"]},{"id":3,"type":"VAELoader","pos":[19.118040084838867,334.3971252441406],"size":[330,60],"flags":{},"order":1,"mode":0,"inputs":[],"outputs":[{"name":"VAE","type":"VAE","slot_index":0,"links":[7]}],"properties":{"cnr_id":"comfy-core","ver":"0.3.59","Node name for S&R":"VAELoader","models":[{"name":"qwen_image_vae.safetensors","url":"file://vae/qwen_image_vae.safetensors","directory":"vae"}]},"widgets_values":["qwen_image_vae.safetensors"]},{"id":6,"type":"EmptySD3LatentImage","pos":[395.3247375488281,447.0367431640625],"size":[211.43600463867188,106],"flags":{},"order":2,"mode":0,"inputs":[],"outputs":[{"name":"LATENT","type":"LATENT","slot_index":0,"links":[6]}],"properties":{"cnr_id":"comfy-core","ver":"0.3.59","Node name for S&R":"EmptySD3LatentImage"},"widgets_values":[1024,768,1]},{"id":9,"type":"VAEDecode","pos":[621.8590087890625,450.3192443847656],"size":[140,100],"flags":{},"order":8,"mode":0,"inputs":[{"name":"samples","type":"LATENT","link":9},{"name":"vae","type":"VAE","link":7}],"outputs":[{"name":"IMAGE","type":"IMAGE","slot_index":0,"links":[10]}],"properties":{"cnr_id":"comfy-core","ver":"0.3.59","Node name for S&R":"VAEDecode"},"widgets_values":[]},{"id":8,"type":"KSampler","pos":[784.2811279296875,128.77174377441406],"size":[300,262],"flags":{},"order":7,"mode":0,"inputs":[{"name":"model","type":"MODEL","link":8},{"name":"positive","type":"CONDITIONING","link":3},{"name":"negative","type":"CONDITIONING","link":4},{"name":"latent_image","type":"LATENT","link":6}],"outputs":[{"name":"LATENT","type":"LATENT","slot_index":0,"links":[9]}],"properties":{"cnr_id":"comfy-core","ver":"0.3.59","Node name for S&R":"KSampler"},"widgets_values":[570849357276640,"randomize",10,1,"res_multistep","simple",1]},{"id":7,"type":"ModelSamplingAuraFlow","pos":[784.8612670898438,26.388885498046875],"size":[300,58],"flags":{},"order":6,"mode":0,"inputs":[{"name":"model","type":"MODEL","link":11}],"outputs":[{"name":"MODEL","type":"MODEL","links":[8]}],"properties":{"cnr_id":"comfy-core","ver":"0.3.59","Node name for S&R":"ModelSamplingAuraFlow"},"widgets_values":[3]},{"id":10,"type":"SaveImage","pos":[1102.9210205078125,30.788015365600586],"size":[441.4784240722656,491.4784240722656],"flags":{},"order":9,"mode":0,"inputs":[{"name":"images","type":"IMAGE","link":10}],"outputs":[],"properties":{"cnr_id":"comfy-core","ver":"0.3.59"},"widgets_values":["QwenImage/ComfyUI"],"color":"#233","bgcolor":"#355"},{"id":11,"type":"UnetLoaderGGUF","pos":[21.36621856689453,18.614456176757812],"size":[335.11993408203125,86.15998840332031],"flags":{},"order":3,"mode":0,"inputs":[],"outputs":[{"name":"MODEL","type":"MODEL","links":[11]}],"properties":{"cnr_id":"comfyui-gguf","ver":"1.1.4","Node name for S&R":"UnetLoaderGGUF"},"widgets_values":["Qwen_Image-Q3_K_S.gguf"]},{"id":5,"type":"CLIPTextEncode","pos":[385.74078369140625,234.67601013183594],"size":[375.2783508300781,160],"flags":{},"order":5,"mode":0,"inputs":[{"name":"clip","type":"CLIP","link":2}],"outputs":[{"name":"CONDITIONING","type":"CONDITIONING","slot_index":0,"links":[4]}],"title":"CLIP Text Encode (Negative)","properties":{"cnr_id":"comfy-core","ver":"0.3.59","Node name for S&R":"CLIPTextEncode"},"widgets_values":["畸形, 模糊, 低品質, 異常, 偽影"],"color":"#322","bgcolor":"#533"},{"id":4,"type":"CLIPTextEncode","pos":[385.7408752441406,24.839426040649414],"size":[373.1487731933594,161.06478881835938],"flags":{},"order":4,"mode":0,"inputs":[{"name":"clip","type":"CLIP","link":1}],"outputs":[{"name":"CONDITIONING","type":"CONDITIONING","slot_index":0,"links":[3]}],"title":"CLIP Text Encode (Positive)","properties":{"cnr_id":"comfy-core","ver":"0.3.59","Node name for S&R":"CLIPTextEncode"},"widgets_values":["一隻可愛的柴犬坐在公園長椅上，日系插畫風格"],"color":"#232","bgcolor":"#353"}],"links":[[1,2,0,4,0,"CLIP"],[2,2,0,5,0,"CLIP"],[3,4,0,8,1,"CONDITIONING"],[4,5,0,8,2,"CONDITIONING"],[6,6,0,8,3,"LATENT"],[7,3,0,9,1,"VAE"],[8,7,0,8,0,"MODEL"],[9,8,0,9,0,"LATENT"],[10,9,0,10,0,"IMAGE"],[11,11,0,7,0,"MODEL"]],"groups":[],"config":{},"extra":{"ds":{"scale":0.9090909090909091,"offset":[41.05623298094109,71.10242663728752]},"frontendVersion":"1.25.11","VHS_latentpreview":false,"VHS_latentpreviewrate":0,"VHS_MetadataImage":true,"VHS_KeepIntermediate":true},"version":0.4}

▲ ComfyUI 的工作流截圖。該工作流就是在 12GB VRAM 上以 ComfyUI 結合 Qwen-Image（Q3 量化）完成本地圖片生成的完整做法。若你想擴充到更複雜的主題或風格，僅需替換提示詞或提高步數即可；若需要更大的解析度，建議先確保顯存充足，再逐步提高影像尺寸或使用的模型等級，並觀察取樣時間與畫質之間的平衡。

中文提示詞示範與生成成果

以下九張示範皆由相同流程生成，提示詞完全以中文撰寫。這次用到的提示詞如下：

1. 一隻可愛的柴犬坐在公園長椅上，日系插畫風格
2. 一隻胖胖的橘貓蜷縮在窗邊的坐墊上，日系插畫風格
3. 一個穿著水手服的少女手捧著氣球，站在海邊的木棧道上，日系插畫風格
4. 一隻小狐狸趴在滿是落葉的石階上，抬頭望著天空，日系插畫風格
5. 一位戴著圓眼鏡的少年正在咖啡廳裡看書，桌上有一杯熱咖啡，日系插畫風格
6. 一隻小企鵝在雪地上玩耍，旁邊有一個小雪人，日系插畫風格
7. 一個少女撐著透明雨傘走在櫻花雨下的街道上，日系插畫風格
8. 一隻貓頭鷹站在圖書館的書架上，背景是溫暖的黃燈光，日系插畫風格
9. 一對情侶並肩坐在河畔，看著夕陽的倒影灑落在水面上，日系插畫風格

▲ 一隻可愛的柴犬坐在公園長椅上，日系插畫風格。背景綠意與陽光斑駁讓畫面很有午後的輕鬆感，木椅的木紋細節與柴犬的臉部表情都相當穩定，邊緣過渡自然，顏色乾淨不堆疊。

▲ 一隻胖胖的橘貓蜷縮在窗邊的坐墊上，日系插畫風格。窗外的柔光製造溫暖的高光區域，毛髮層次與花紋的銜接很平滑，坐墊的布料與陰影也保持乾淨的筆觸，整體極具療癒感。

▲ 一個穿著水手服的少女手捧著氣球，站在海邊木棧道上，日系插畫風格。海天線明確、雲朵鬆軟，裙擺與髮絲因海風微動，構圖留白剛好，色彩清爽而且具有夏日氣息。

▲ 一隻小狐狸趴在滿是落葉的石階上，抬頭望著天空，日系插畫風格。金黃與橘紅的落葉與狐狸毛色相互呼應，階梯的粗糙質地與柔焦背景讓季節感十分明確。

▲ 一位戴著圓眼鏡的少年正在咖啡廳裡看書，桌上有一杯熱咖啡，日系插畫風格。景深將背景柔化，把注意力集中在人物與書本，杯盤與紙張的材質處理細膩，氛圍安靜而專注。

▲ 一隻小企鵝在雪地上玩耍，旁邊有一個小雪人，日系插畫風格。雪花粒子與地面反光處理俏皮清透，角色表情靈動可愛，對比度恰到好處，整體畫面乾淨且無明顯噪點。

▲ 一個少女撐著透明雨傘走在櫻花雨下的街道上，日系插畫風格。粉色花瓣與路面濕潤的高光交織，人物邊緣光柔和，從構圖到氣氛都呈現輕盈與浪漫。

▲ 一隻貓頭鷹站在圖書館的書架上，背景是溫暖的黃燈光，日系插畫風格。羽毛紋理與眼神聚焦清晰，暖黃光影讓空間顯得沉靜，書脊排布提供了良好的規律感。

▲ 一對情侶並肩坐在河畔，看著夕陽的倒影灑落在水面上，日系插畫風格。逆光邊緣光與水面反射自然，前景草叢作為框景增添層次，整體情緒柔和而細膩。

▲ 九張示範的縮圖拼接總覽，日系插畫風格在光影、邊緣與色塊的穩定度表現都相當不錯，搭配 Q3 量化基本能兼顧速度與細節。