最近經常看到網上說DeepSeek R1有多麼強大，於是便想在ComfyUI上使用看看是否能透過它將提示詞優化，但我的deepseek.com很早就註冊了，贈送的餘額已經失效，在together.xyz上看到它也有提供DeepSeek R1(雖然不知道模型規模是否一致)，所以就決定使用它提供的DeepSeek R1，以前都用comfyui_llm_party ☁️API LLM general link這個節點，但它不能用together.xyz的樣子，替換了API_URL也是呈現連線錯誤，原因不明~
於是老樣子叫了AI幫我寫這個ComfyUI Custom Nodes，讓它可以呼叫together.xyz
節點的使用，輸入api_key跟要用的model，系統提示詞、用戶提示詞...
輸出
response_text:完整的回傳
filtered_text:將<think></think>的部分給過濾掉，只留結果。
think_content:顯示思考過程。(呈現<think></think>的部分)
實測下來，發現它的思考過程挺不錯的。但實際回應的內容，好像沒有想像中的好，我測試的範例是下圖，一隻熊面前有顆石頭，想透過LTX Video來產生影片。
下面是直接從官網的回應(思考過程)
實際產生的prompt
the large, muscular brown bear rises onto its hind legs with deliberate power, thick shoulder muscles rippling beneath dense cinnamon-brown fur streaked with amber highlights, its massive front paws pressing against a weathered granite boulder twice the size of its torso. coarse facial fur bristles around intense amber eyes locked forward with primal focus, sunlight catching individual guard hairs as the beast exhales visible vapor into the crisp mountain air. jagged slate-gray rocks form an unstable slope beneath its splayed hind claws, gravel cascading downward as shifting weight crunches stone - the bear's every tendon strains visibly through its fur while pushing the 800-pound rock toward the precipice. behind this primal struggle, mist coils around distant snow-capped peaks under a brooding steel-blue sky, wind whispering through sparse alpine grasses clinging to cracks in the granite battlefield. the camera glides laterally at knee height on a steady dolly track, wide 24mm lens exaggerating the bear's monumental scale against the collapsing rockfield, capturing synchronized details: soil churning under sliding boulder, individual pebbles suspended mid-fall, droplets of saliva flung from the bear's snarling muzzle. hyper-realistic CGI mimics Arri Alexa footage with subsurface fur scattering golden-hour light, every sweat-matted hair strand reacting to muscular torque as the stone finally breaks free - a slow-motion avalanche of splintering shale erupts where rock meets slope, the bear's triumphant roar echoing through valleys while its still-extended claws glint with mineral dust. scene is captured as photorealistic VFX movie footage in the style of The Revenant's visceral naturalism, rendered with Redshift's volumetric lighting to emphasize raw physicality amidst primordial geological forces.
下面則是使用ComfyUI呼叫的together.xyz上的R1，我覺得它的思考過程挺不錯的，有充分的理解需求。但就是回應的內容，好像沒思考的強大~
也不知道是不是LTX的語言訓練集的關係，用R1產生出來的提示詞，呈現的畫面就挺超出預期的，用Meta Llama 或者 Open AI 就好一些。
這是gpt-4o-mini的回應
這是gpt-4o的回應
以後再想想能怎麼運用它，有哪些適合的場景。(或許嘗試將它的推理過程跟原本的需求詞融合提交給gpt，不知道會不會有更好的結果)
文章轉載或引用，請先告知並保留原文出處與連結！！(單純分享或非營利的只需保留原文出處，不用告知)
原文連結:
https://blog.aidec.tw/post/comfyui-deepseek-r1-llm-together
若有業務合作需求，可寫信至： opweb666@gmail.com
創業、網站經營相關內容未來將發布在 小易創業筆記