Zhenzhen_Mj

Midjourney is the gold standard for artistic image quality and, until this pack, a pain to use from ComfyUI - it lives in Discord, takes a subscription, and has zero interest in your graph. Comfly_Mj is the flagship node that changes that: it's a text-to-image Midjourney node with the whole parameter deck exposed. Speed modes, model versions back to v 5.1, the niji anime line, chaos, stylize, style/character/object references, aspect ratio - it's all here as inputs instead of you hand-typing --ar 16:9 --s 250 into a chat box.

It's the reason people install this pack. If you've ever wanted Midjourney's look without leaving your normal workflow, this is the node. Paid per call through the Comfly proxy - your key at ai.comfly.chat, cloud inference, GPU idle.

How it works

The node assembles a real Midjourney prompt string from your inputs - appending --v 6.1, --ar 1:1, --s 250, --cref <url> and friends exactly as you'd type them in Discord - then submits it to the proxy's /mj/submit/imagine endpoint on the route that matches your speed choice (turbo, fast, or relax). It polls until the 4-up grid is ready and hands back the image, the prompt text, and a taskId - that taskId is the key to the next step, because it's what you feed into Comfly_Mju to upscale one of the four grid cells.

The inputs that matter

• text - your prompt. The README's audience is Chinese-first, and there's a text_en field alongside it: if you fill text_en, that's what MJ actually gets (handy when your main prompt is in Chinese and you've translated it once).
• speed - turbo mode / fast mode / relax mode, default fast. Turbo is quicker and pricier; relax is cheap and slow. This sets the proxy route.
• model_version - v 7, v 6.1 (default), v 6.0, v 5.2, v 5.1, and the niji anime line (7/6/5/4). v 7 is the current big architecture (draft-mode fast generation, style consistency), v 6.1 is the beloved mature default.
• ar - aspect ratio, free string (1:1, 16:9, ...).
• no - what to exclude. c - chaos (0–100). s - stylize (0–1000). iw - image weight (0–2, for image references). sw / cw / sv - style/color weight and style-variation tier. r - repeat (1–40).
• sref / cref / oref - style, character, and object references. These take URLs - wire in the url output from Comfly_upload, the pack's upload node, to use one of your own images. Default "none" means off.
• positive - extra prompt text appended to the main prompt.
• video / tile - booleans: video asks for the animation MP4, tile for seamless tiling.
• seed - lock a take.

Outputs: image (IMAGE), text (STRING - the assembled prompt), taskId (STRING - feed to Comfly_Mju).

Install

Manager → search Comfyui_Comfly, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ainewsto/Comfyui_Comfly

Restart. Standard requirements.txt (aiohttp, requests, Pillow, numpy + existing torch/transformers), no model downloads. Key from ai.comfly.chat in Comflyapi.json or the node's api_key.

Where people get burned

The classic new-user miss: you generate, get a grid, and wonder why the image is low-res - because MJ returns a 4-up grid and you never ran Comfly_Mju to upscale the cell you actually want. That's the two-node dance this pack is built around. Also, sref/cref need hosted URLs, not local images - skip Comfly_upload and the reference silently no-ops. And relax mode is slow; don't file a bug report when a relax-mode job takes five minutes. It's a third-party proxy in front of MJ's own Discord-based system, so treat the taskId history as ephemeral and save your finals.