Jib Mix Qwen · Text to Image

J1B is the person behind the Jib Mix Realistic XL merges on Civitai - the model page is 194768 if you want to check the lineage - and Jib Mix Qwen is the same instinct applied to Alibaba's 20B Qwen-Image base: a community merge tuned for natural faces and skin rather than for prompt gymnastics. On SpicyAPI it runs on their hardware for $0.024 an image, flat. No 24GB card, no fp8 conversion, no twelve-minute queue while the text encoder loads.

That's the actual pitch. On a local box, Qwen-Image is a VRAM hog that the community's honest summary describes as "way heavier than the others" - great with LoRAs, underrated as a refiner, but not the daily driver its quality deserves. At 2.4 cents a frame with no setup, the tradeoff flips: you're paying to skip the hardware problem entirely.

What the node does

One prompt in, one image out, about six seconds of server time. Nine aspect ratios from 1:1 up to 21:9 and 9:21 (that tall/strip pair is unusually generous), with output at up to 1536 pixels on the long side. No content filter, which is worth stating plainly because it's the thing the local Qwen family is known for - Alibaba shipped the whole line Apache 2.0 and does not police fine-tunes, unlike some of the Flux-era licences.

The 1536 cap is the honest caveat. By 2026 standards that's small. Treat this as a generator for a composition you're going to finish somewhere else - feed the output into your local upscale or a detail pass rather than expecting a final 4K file out of it.

Inputs worth knowing

• prompt - describe the person, the clothing, the place, the light and the camera feel. Quote anything you want rendered as text in the image; Qwen-Image was built as a text-rendering model and the merge kept that ability.
• aspect_ratio - 1:1 by default, then 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9, 9:21.
• output_format - jpeg (default), png or webp. Ask for PNG if you're going to run more edits on it; JPEG is fine for review passes and smaller downloads.
• seed - an integer, and the tooltip is the important part: same seed and settings give the same result, and every run is billed. The node defaults to fixed control so re-queueing the same graph doesn't quietly charge you twice. Flip control after generate to randomize when you want a genuinely new sample.

Outputs are images (IMAGE) and task id (STRING - the handle for the run in the SpicyAPI console, and the thing support asks for).

Two things are missing on purpose: there's no negative prompt and no steps/CFG/sampler controls. Everything sampling-related is the platform's business, not a widget. If you're used to herding a local sampler, that absence takes an afternoon to get used to, and it's also why the results are consistent between your machine and mine.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/spicyapi-ai/comfyui-spicyapi

Restart ComfyUI, open Settings > SpicyAPI, paste a key from spicyapi.ai/console/keys, restart once more. Manager search: "SpicyAPI". On a server, set SPICY_API_KEY. The pack declares zero extra dependencies - Python 3.10+, ComfyUI's V3 node API, and that's it. No weights to download, which is a strange and pleasant sentence to write about a model this size.

The community's read on the merge

Worth knowing before you judge the output: J1B's Qwen merge drew the usual "smells like Flux, should've been a LoRA" critique when the v5 showcase landed - the point being that these merges pull realism in a particular direction that no longer reads as vanilla Qwen. People who like it really like it; one commenter in that thread said they'd had more luck with the merge than with any LoRA they could find. If the faces look photographed rather than rendered, that's the model working as advertised, not a bug.

Practical traps

Billed per run, no free preview. Seed-fixed defaults mean the repeat is cheap only in the sense that it costs nothing to re-view. Changing the prompt means paying again. Batch-thinking is expensive thinking.

No filter cuts both ways. There's no refusal to fight, but there's also no moderation net, so "no content filter" means you're the one checking what you publish, and the destination platform's rules still apply.

Errors are the pack's standard set: No SpicyAPI key is set if you skipped the settings step, 40310 for an unverified email, 40004 when that exact combination isn't served. Keep the task id output connected to a text display and you'll have the exact handle for anything that goes wrong.