Midjourney Generate (Direct via LegNext)

Midjourney Generate (Direct via LegNext) is the one node in the ComfyUI-Midjourney-Direct pack, and the name deserves a footnote: it doesn't talk to Midjourney directly, and it isn't free. Midjourney has no official public API, so this node calls LegNext.ai, a pool-based proxy that rents real Midjourney accounts and bills you per image - roughly $0.08 for a fast render, no monthly minimum. What you get out is genuine Midjourney output (v6 through v7, niji-6, and the v8 line with a Pro sub) as a normal IMAGE tensor in your graph. No Discord, no subscription juggling, no moderation roulette.

Why bother when Flux and friends run locally? Because Midjourney still has a look. Our KB's closed-source-models essay calls it "the defining benchmark the open-source community measures against" - consistent artistic finish that keeps pulling people back even as the open gap shrinks. This node is for the person who wants that look as a building block - drop the 4-grid into an upscaler, a face-restore pass, or an img2img chain. It is not for people who just want MJ pictures; Midjourney's own web app is easier and cheaper at volume. This is a pipeline tool, not a gateway.

How it works

It's a thin REST client, and the source is refreshingly honest about it. The node assembles your prompt into a plain Midjourney flag string - --v 7 --ar 2:3 --style raw and friends - POSTs it to LegNext's diffusion endpoint with your key in an x-api-key header, gets back a job_id, then polls the job endpoint every few seconds until it's done. Then it downloads the result and hands you the familiar 4-grid as a single composite IMAGE, typically 1024×1024 with four 512×512 tiles. All the actual MJ inference happens on LegNext's servers; your GPU just waits, usually 45–90 seconds.

The inputs that matter

Only four are required, and you'll set three of them:

• prompt - your imagine prompt, multiline. Don't stuff flags in here; wire them as structured inputs below.
• version - defaults to v7 (default), which is right for the standard pay-as-you-go tier. v8.1 / v8 / v8 hd need a LegNext Pro subscription on top. niji-6 is the anime/East-Asian aesthetic model.
• aspect_ratio - 1:1 through 21:9, default 2:3.
• seed - 0 is random; a fixed seed plus the same prompt and params gives consistent output.

The optional list is where the Midjourney muscle lives: style_raw (on by default - reduces MJ's heavy default stylization, the AKURATE-realism move), stylize (0–1000, lower is more prompt-faithful), weird (0–3000, unconventional aesthetics), chaos (0–100, variety across the grid tiles), quality (0.25–4, higher costs more), turbo (4× faster, 2× price), draft (cheap fast preview for ideation), image_weight (only matters if your prompt contains an image URL), negative (the --no list), and tile. Leave poll_interval_seconds (3) and timeout_seconds (300) alone until something times out.

Outputs

• image (IMAGE) - the 4-grid composite. Split or crop it downstream; individual-tile upscaling is a v0.2 feature that doesn't exist yet.
• cost_info (STRING) - points consumed, approximate USD, elapsed time, and the actual seed. A nice reality check on how much ideation costs.
• cache_key (STRING) - a deterministic hash of prompt + flags + seed. Wire it into the same author's ComfyUI-API-Optimizer Hash Vault and identical re-runs return instantly at $0. This is the pack's smartest feature - reuse it.

Install

ComfyUI Manager (search "ComfyUI Midjourney Direct") or the manual way:

cd ComfyUI/custom_nodes/
git clone https://github.com/jeremieLouvaert/ComfyUI-Midjourney-Direct.git
pip install -r ComfyUI-Midjourney-Direct/requirements.txt

Restart ComfyUI and it appears under Midjourney Direct. Dependencies are just requests, Pillow, numpy - no model downloads, no heavy wheels, nothing to hurt your VRAM.

API key setup

Grab a key at legnext.ai. The node looks for it in order: the api_key input → the LEGNEXT_API_KEY environment variable → a legnext_api_key.txt file in your ComfyUI root. The file is the recommended option - set it once and forget it.

Where people get burned

• "No LegNext API key found" - you forgot all three places. Do the file, it's the least fiddly.
• Flag rejections. LegNext occasionally rejects flag combos Midjourney itself accepts. Job fails with a flag-ish error? Strip the exotic stuff (weird, chaos, odd quality values) and retry.
• Timeout on a busy pool. Default cap is 300s; generations normally land in 45–90s, but a backed-up pool can blow past that. Bump timeout_seconds if you keep hitting the wall.
• Low-res tiles. You're getting four 512×512 images in one grid, so plan an upscale pass.

This is v0.1, remember - imagine-only, no U1–U4 upscaling, no V1–V4 variation, no reroll or zoom. If that matters, it isn't there yet. What is there works, costs nickels, and drops Midjourney into your graph like it always belonged.