Jib Mix Qwen LoRA · Text to Image

This is the version of Jib Mix Qwen that takes LoRAs - up to three, each with its own strength, for $0.03 an image. Same photoreal merge of Alibaba's 20B Qwen-Image underneath, same nine aspect ratios, same 1536-pixel ceiling, same no-content-filter setup. What changes is the loras input, and that one socket is why this node exists as a separate entry rather than a checkbox on the plain one.

Why route LoRAs through an API at all

Because Qwen-Image + LoRA is exactly the combination that's fiddly locally. J1B said it himself in the thread about his own merge: LoRA compatibility on Qwen is rough in practice, "poor quality and line artefacts as well as poor adherence to the lora, especially with the fp8 version that most people with under 24GB of Vram use." That's a real complaint about a real local constraint, and it's the strongest argument for this node. SpicyAPI runs the LoRA against the base model on their hardware, at full precision, so the failure mode where your LoRA half-works and half-smears line artefacts on a consumer card isn't your problem. You pay 0.6 cents over the non-LoRA node per image to skip it.

The other reason: the community's whole Qwen story is "it responds insanely well to LoRAs." If you already have a Qwen-Image LoRA you like, this is the shortest path from file to image on a laptop.

Wiring your LoRA in

You don't paste a path into this node - the loras input takes SpicyAPI LoRA nodes, chained. Each one holds a direct download URL to a .safetensors file and a strength:

• Chain up to three LoRA nodes into loras; strengths run 0 to 4, which is a wider range than a local ComfyUI LoRA loader gives you, so you can genuinely overcook a look if that's what you want.
• The link must open without a login. A Hugging Face URL that bounces to a sign-in page will fail; use the direct file URL, and for gated repos check that anonymous access works before blaming the node.
• The LoRA must be trained for Qwen-Image or Qwen-Image-2512. This is the one that bites. A genuinely excellent SDXL or Flux LoRA is simply the wrong shape here; the API will not magically adapt it.
• Trigger words go first. The node's own tooltip says it: if your LoRA has a trigger phrase, put it at the front of the prompt.

Start with one LoRA at 0.8–1.0, look at the result, then add a second on a different axis (style + subject) rather than two versions of the same idea.

The rest of the node

Everything else matches the plain Jib Mix Qwen node. prompt describes the subject and light; aspect_ratio offers 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9, 9:21 (default 1:1); output_format is jpeg, png or webp; seed is an integer whose tooltip reminds you that seeds start fixed and every run is billed. Outputs are images (IMAGE) and task id (STRING) - keep the task id visible, it's the console handle for a bad run.

No negative prompt, no steps, no CFG. Sampling is the platform's business.

Install

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

Restart, add your key in Settings > SpicyAPI (from spicyapi.ai/console/keys), restart once more so your account's model list loads. Manager search is "SpicyAPI". For a headless or cloud box, SPICY_API_KEY in the environment does the same job. Zero dependencies beyond ComfyUI - no models, no CUDA, nothing to compile.

Once you're in, the pack ships a lora template in the ComfyUI template browser that's wired and ready to run with an Apache-2.0 pixel-art LoRA for Z-Image Turbo, which is the fastest way to see how the chain is supposed to look.

Cost and failure notes

Three cents an image, every image. LoRA iteration is the classic place to burn money, because the useful experiment is variation. Set a max cost per run in the settings and accept that a sweep of 30 images is ninety cents - cheap by local standards, but it's ninety cents every time you reshuffle.

LoRA silently ignored is the most likely disappointment, and it's almost always one of the three rules above: a login-walled URL, a LoRA for the wrong base, or a trigger phrase buried in the middle of the prompt.

Uploads and errors. Images aren't your only traffic - the LoRA files are fetched server-side, so a slow or dead host can stall a run. The usual pack errors apply: No SpicyAPI key is set, 40310 for an unverified email, 40004 if a chosen option combination isn't served, and an HTTPS_PROXY misconfiguration shows up as a message about a proxy or firewall intercepting the request.