LongCat Text to Image

Most image models render text the way a toddler draws a car - recognizably a car, if you squint. LongCat-Image is one of the few open models that writes actual, legible characters, and its superpower is Chinese: signs, posters, product labels, everything comes out readable instead of pseudo-character gibberish. This node is where you type what you want it to draw.

It's a text-to-image node in the classic sense, but the model underneath is a 2026-style architecture: a 6B flow-matching transformer with a Qwen2.5-VL LLM as the text encoder. That changes how you prompt. You talk to it in full sentences like you're describing a photo to a person - 一个年轻女性站在雨中的霓虹灯招牌前，回眸微笑 - not in a pile of comma-separated tags. The old SD habits actively hurt here: masterpiece, best quality, 8k does nothing on an LLM encoder, and the negative prompt barely participates in the result. Keep it empty unless you're fighting something specific.

How it works

You feed it the longcat_pipeline from the LongCat Model Loader, plus your prompt. It passes the text through the Qwen encoder, runs the flow-matching denoise, and hands back a normal IMAGE tensor you can preview, save, or push into an upscaler. Two LongCat-specific switches ride along: enable_cfg_renorm and enable_prompt_rewrite, both defaulting to true. The model was trained with these on - this is not the place to experiment. Leave them alone.

The inputs that matter

• prompt - natural language, and here's the rule that matters most: any text you want rendered into the image goes in quotes. "你好世界" or "coffee shop" - single or double quotes, English or Chinese. The model uses a special character-level encoding for quoted content; skip the quotes and your text comes out mangled. This is the model's own documented behavior and the #1 beginner mistake.
• steps - 50 for the base model, 28 for Dev. People do report decent results at 15–20 when iterating; quality just softens. You're not locked to 50.
• guidance_scale (default 4.5) - this is a guidance-distilled flow model, so stop thinking in SDXL CFG 7. If your output looks harsh and blown-out, you pushed the slider up; dial it back instead.
• width / height - multiples of 64, between 64 and 4096. The 1344×768 default is a wide, banner-ish ratio and a fine place to start.
• seed - for reproducing a result you liked. Set it, note it, done.
• batch_size - 1 to 8 images in one go. Mostly useful once you've got a prompt you trust; batch of 8 at 50 steps is a long wait.

One output: IMAGE. That's the whole deal.

The pair gotcha

Wire the Edit model into this node and it refuses outright: "Use LongCat Image Edit node for edit models." The loader decides which pipeline you get by reading the folder name, and this node only accepts a generation model (LongCat-Image or LongCat-Image-Dev). If you see that error, you picked the wrong model at the loader, not the wrong settings here.

Troubleshooting

• Garbled text → you forgot the quotes. Rewrap the text in "…" and rerun.
• Harsh, oversaturated output → guidance too high for the steps. Try guidance_scale around 3.5–4.5 with 20+ steps.
• OOM → you turned off CPU offload at the loader. Turn it back on; it's the difference between running in ~17GB and needing 24GB+.

The wrapper itself is young and low-star, but the model does what it claims, and the Chinese text rendering genuinely is the differentiator - there aren't many open weights where you can generate a readable menu board on the first try.