JibMix Qwen Image

JibMix Qwen Image is a text-to-image node that runs a portrait-optimized variant of Qwen Image through the WaveSpeed API, at the same $0.02-per-image price as the standard Qwen Image node. Same interface, same bilingual (Chinese/English) prompt support, same 256–1536px range - the "JibMix" part is the model recipe, which is tuned toward people and portraits rather than generic scenes.

Honest framing: "portrait-optimized" is a marketing-ish claim you should verify against your own taste. Some people find the JibMix variant noticeably better at faces and skin, others can't tell the difference and just use plain Qwen Image. It costs the same as the default, so if you're doing people-heavy generation it's worth an A/B test on your own subject matter. The Qwen family itself is Alibaba's - the same lab behind the open-weights Qwen-Image that the local community has been training LoRAs on since mid-2025 - so the underlying model lineage is solid, just served commercially here.

How it works

The node sends your prompt to WaveSpeed's JibMix Qwen endpoint (a size of width*height, an output format, and a sync mode flag), waits for the render, and converts the returned URL into a real ComfyUI IMAGE tensor. Because the output is a tensor, it plugs straight into your existing image chain - no latent decode step, no sampler. There's a seed widget, but as with most of the pack's WaveSpeed image nodes it's cache control rather than a true API seed.

The inputs that matter

• prompt - the description, in Chinese or English. Qwen models handle both well.
• width / height - 256 to 1536, step 8, default 1024×1024.
• output_format - jpeg (default), png, or webp.
• enable_sync_mode - wait for completion before returning. Off by default, which lets the node return faster and poll in the background.
• enable_base64_output - return base64-encoded output instead of a URL. Only matters if you're scripting against the API.

The optional client input wires to a WaveSpeed API Client node; unconnected, the key comes from Settings.

Installing it

This node is part of the ERPK Collection (eRepublik-Labs/comfyui-nodes-erpk), the monorepo that bundles WaveSpeed, Grok, Claude, Gemini, and OpenAI nodes. ComfyUI Manager: search erpk → ERPK Custom Nodes → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk
pip install -r requirements.txt

You'll need a WaveSpeed API key (Settings → ERPK → API Keys). The WaveSpeed dependency footprint is just pydantic/requests/Pillow - no models to download.

Common issues

• "Prompt is required" - blank prompts are rejected before they hit the API.
• No image in the output - if generation fails server-side, check your WaveSpeed credit balance; this bills per image.
• Faces still off? - the model does what it does; try bumping to 1536px and being specific about lighting and framing in the prompt. You're paying per attempt, so budget your iterations.
• Base64 output surprises - if you flip enable_base64_output on, expect a base64 blob rather than a normal URL; leave it off unless you have a reason.