Tensor.Art Generation Guide: Models, Prompts and Settings

Tensor.Art (tensor.art) is a community platform for Stable Diffusion and Flux style models with generation built in. You browse checkpoints and LoRAs uploaded by other users, generate with them in the browser on a credit system, and publish results back with their settings attached. If you have used Automatic1111 or Civitai’s generator the controls will feel familiar; if you have not, this guide explains what each one does and where to start.

Finding a model to start from

Models on Tensor.Art fall into the usual categories: checkpoints (the full model), LoRAs (style, character or concept add-ons) and embeddings, plus workflows for people who want node graphs. Each model page shows the base model family, example images with their generation data, the creator’s notes and any trigger words. Filter by base model first, because a LoRA only works with checkpoints from the same family (for example SDXL LoRAs with SDXL checkpoints).

The quickest way to a good first image is to open an example you like, use its remix option to load the full prompt and settings into the generator, and run it unchanged. If the result matches, your setup is right and you can start altering things with confidence.

Writing the prompt

Most checkpoints on the site are SDXL-era or similar, and they respond to the classic comma-separated style. A few habits make a real difference:

• Subject first. The first few words carry the most weight. "portrait of an elderly fisherman" beats "a photo, high quality, of a fisherman".
• Trigger words up front. If a LoRA lists a trigger word, place it at or near the start, spelt exactly as given.
• Then setting, light, lens. "on a harbour wall at dawn, overcast light, 50mm lens". These three categories cover most of what people forget.
• Style last and sparingly. One or two style words. Long strings of "masterpiece, best quality, 8k" help little on modern checkpoints and can push everything towards the same glossy look.
• Weighting. Many interfaces, including this one in its current form, accept parentheses to emphasise a word, such as (red scarf:1.2). Use it on one or two elements, not the whole prompt.
• Flux-style checkpoints are different. They prefer plain sentences, ignore most quality words and usually do not use a negative prompt. Check which family you are on.

The negative prompt lists what to avoid. A short, general list works for almost everything: "blurry, low quality, deformed hands, extra fingers, text, watermark". Add specifics only when a problem keeps appearing.

The settings, and what to set first

1. Checkpoint. Decide the family and the base look.
2. LoRAs and weights. One or two, at the creator’s recommended weight or around 0.7. Lower the weight if the LoRA dominates every image.
3. Size. Use the family’s native resolution (about 1024 on the long side for SDXL and Flux). Odd sizes far from native produce doubled subjects.
4. Sampler and steps. Keep the creator’s recommended sampler; 25 to 30 steps is a reliable range. Beyond about 40 you pay more credits for no visible gain.
5. CFG scale. Prompt adherence. SDXL checkpoints usually like 5 to 7. Higher values oversaturate; lower values go soft.
6. Seed. Random while exploring, fixed while refining. A fixed seed is the only way to know whether a change in output came from your edit or from chance.
7. Batch count. Four images per run while exploring, one while refining.

You may also see options for a VAE, clip skip, hires fix or an upscale pass, and a ControlNet section for guiding composition with a reference image. Leave the first two at the model’s recommendation. Treat the upscale pass as a final step on a chosen image rather than something you run on every draft.

Settings at a glance

For the theory behind these controls see Stable Diffusion settings explained. For workflows and model training on the same site, see Tensor.Art workflows and models.

Credits and keeping records

Generation is paid in credits, and the cost rises with size, steps, batch count and some model choices. The interface shows the cost before you run, and there is usually a daily allowance on free accounts; the amounts change, so check the platform’s current pricing page. Spend small while exploring and large only on a proven prompt.

Each image you make keeps its prompt and settings on the site, but a downloaded file may not carry them. When an image matters, copy the checkpoint name and version, LoRA list with weights, prompt, negative prompt, sampler, steps, CFG, size and seed into your own notes. If you can reproduce the image from the notes alone, the record is good enough.

Try it

Checkpoint: a stylised illustration SDXL checkpoint of your choice
LoRA: none for the first run
Prompt: a lighthouse keeper reading by lamplight in a small round room, rain on the window, warm interior against cold blue night, storybook illustration, soft textured shading
Negative prompt: blurry, low quality, deformed hands, text, watermark
Size: 832x1216. Steps: 28. CFG: 6. Seed: random, batch of 4

Choose the best image, fix its seed, then change only the CFG to 4 and then 8. The three results show the model’s range on that one control, which teaches you more than any written description.