Z-Image Generate

Z-Image without the graph

If you just want a Z-Image picture and can't be bothered to wire up the native ComfyUI Z-Image loader, sampler, VAE and text encoder, this node is the shortcut. It's one box on the canvas: type a prompt, pick a resolution, hit go. Behind the curtain it loads Alibaba's Z-Image Base (Tongyi-MAI/Z-Image) through Hugging Face Diffusers, runs the whole pipeline internally, and hands you a plain IMAGE tensor you can preview or save like any other.

The catch is in the name: this is a wrapper around the whole model, not a native ComfyUI implementation. You trade fine-grained control for "it just works." For a quick idea, that's often the right trade. For anything serious you'll outgrow it fast - it has no sampler selection, no LoRA loading, no ControlNet, no upscaling.

What you're actually running

Default model_id is Tongyi-MAI/Z-Image - that's the Base model, the non-distilled 6.15B checkpoint Alibaba shipped in January 2026, not the faster Turbo. The difference matters. Base wants CFG 3–5 and 25–50 steps; Turbo is guidance-distilled, wants CFG 1 and 8–9 steps. The node's defaults (25 steps, guidance 4.5) are right in the Base sweet spot, which suggests the author actually tuned them for Base. If you switch model_id to Tongyi-MAI/Z-Image-Turbo, keep those defaults and you'll get burned, overcooked images. Also worth knowing: because the text encoder is Qwen3-4B, Chinese prompts land noticeably better than English ones.

Mechanically the node is simple and readable: DiffusionPipeline.from_pretrained(model_id, dtype=...), then one pipe(...) call inside torch.inference_mode(). The pipeline is cached in a module-level dict keyed on (model_id, device, dtype), so the second generation doesn't re-load the model. There's no ComfyUI-native Z-Image quirks here - which is a genuine upside, because it also means the infamous Sage Attention corruption on Base never applies; diffusers doesn't use ComfyUI's attention patches.

The inputs that matter

Most of the nine fields are self-explanatory. The ones to actually touch:

• prompt - natural language, not tags. Z-Image is a prose model.
• width / height - 256 to 2048, step 64. Stay in the 1–2MP native range (1024² or 1216×832); past ~2MP quality degrades.
• steps / guidance_scale - 25/4.5 defaults are fine for Base. Add steps if a prompt fights you.
• negative_prompt - optional, and it works on Base (it's meaningless on Turbo).
• clear_cache - leave off for repeated generations; flip it on if you want VRAM back after one shot.

The output is a single IMAGE, wired straight into a PreviewImage or SaveImage node. That's it - one image per run, no batch.

Install

The README's own clone command has a placeholder in it, so copy with your eyes open:

cd ComfyUI/custom_nodes
git clone https://github.com/suravaya113/SKIT_ComfyUI-ZImage-Generate
pip install -r SKIT_ComfyUI-ZImage-Generate/requirements.txt

Then restart ComfyUI. Or just search "SKIT_ComfyUI-ZImage-Generate" in ComfyUI Manager. Two real gotchas: the requirements pin diffusers straight from GitHub main (diffusers @ git+https://github.com/huggingface/diffusers.git) plus transformers==4.57.3, so it can clobber your existing diffusers install - make sure ComfyUI's native Z-Image support still works after. And the first run downloads the whole Base model (plus the 4B Qwen text encoder) from Hugging Face. That's many gigabytes and a long load; a 12GB card will run it, and it'll use ~3–4 minutes per image there.

Troubleshooting

• "ZImageGenerate requires diffusers and related packages" - exactly what it says: you skipped the pip step. Install requirements and restart.
• Slow first generation - it's downloading the model, not hung. Run once, let it finish, and subsequent gens are fast from cache.
• Expecting Turbo, getting slow Base - check model_id. Turbo needs its own settings.
• VRAM pressure with clear_cache on - it unloads after every generation, so the next one re-loads and re-downloads nothing but re-loads from disk. Fine for one-off, wasteful for iteration.

Honest verdict: this is the fastest way to get from zero to a Z-Image picture in ComfyUI, and the defaults are sane. If you stay on Base and one-box simplicity, it'll treat you fine. The moment you want LoRAs, samplers, or speed, graduate to the native nodes - but keep this around for a quick sanity check.