Z-Image Base LoRA · Text to Image

The odd thing about Z-Image Base is that it's the model everyone says to train on and almost nobody wants to generate with. It's slower than Turbo, it needs a proper sampler setup, and it's the better of the two when you care about variety and prompt adherence. So a hosted Base text-to-image node is basically a LoRA sandbox: bring your own weights, skip the 6B checkpoint and the VAE and the text encoder and the node graph that holds them together.

Why Base, again

Quick recap if you skipped the ecosystem reading. Z-Image is 6B parameters, Apache 2.0, from Alibaba's Tongyi-MAI lab, and it landed in late 2025 to a reaction that genuinely surprised people - a 3060 generating usable images in half a minute was a real change. Turbo is the distilled 8-step variant: fast, beautiful, and criticized for producing near-identical images across seeds. Base is the undistilled foundation model, shipped January 2026, and it restored the seed variance and the weird-concept adherence that distillation smooths off.

It's also the model the LoRA ecosystem moved to, because Base LoRAs work across Base and Turbo (Turbo at a higher strength), while Turbo LoRAs are stuck on Turbo. If your fine-tune is a Base one, this is the node that treats it as intended.

The fields, and what's missing

prompt - subject, clothing, place, light, in sentences, with any LoRA trigger phrase first. The tooltip is explicit about the trigger placement, which is a hint that this node's intended use is your own LoRAs rather than generic prompting.

aspect_ratio runs 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9, 9:21 - the full set, including the wide and tall extremes that Base handles reasonably well.

output_format is jpeg, png or webp. seed defaults to fixed, because every run is billed and re-queueing shouldn't silently cost you.

loras is the point of the node: up to three files trained for Z-Image Base, each with a strength. Connect a PopNow LoRA node and set its source - a .safetensors already in ComfyUI/models/loras gets uploaded once and reused for 12 hours, a Hugging Face repo id or page gets imported, a Civitai page gets imported, a direct https:// link is passed straight through. Chain PopNow LoRA nodes to stack them, and mind that Z-Image has historically punished stacking more than one.

What's not here: steps, CFG, sampler, scheduler, resolution tier. There's no 1024-versus-1536 choice, no negative prompt, no LoRA trigger validation. You are renting the pipeline, not the dials. For a lot of people that's the appeal - the KB's sampler sweep on this family ended up at "Euler at 8-9 steps, CFG 1" anyway, so there's not much to tune.

Cost and mechanics

Flat 2.16 credits per image, about two cents at 1 credit = $0.001, LoRAs included. The node builds a request from alibaba/z-image-base-lora/text-to-image's published schema, uploads anything you connected, quotes the job, submits with an idempotency key so retries can't double-bill, polls, then decodes the returned file into an IMAGE batch. It also hands you task id for the API console. Wire the image into Save Image, an upscaler, or back into an edit node.

Install

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

Or install PopNow from ComfyUI Manager and restart. No model files, no dependencies beyond ComfyUI itself, no GPU. Create a key at popnow.ai/api/console/keys, restart, enter it under Settings > PopNow (or POPNOW_API_KEY), restart again so the node list comes from your account.

Where people get burned

Wrong base model, wrong result. A Turbo-trained LoRA fed to a Base model produces mush - that direction of transfer was confirmed broken on release day. If your LoRA came from Civitai, check what it was trained on before blaming the node.

Second: gated imports. A Civitai or Hugging Face file behind a login needs CIVITAI_API_KEY or HF_TOKEN in ComfyUI's environment. The pack deliberately keeps tokens out of widgets, so there's no field to paste one into - set the variable and restart.

Third: expecting photo-restoration from a 2-cent render. Base is good. It is not a magic eraser for a bad prompt.