Tensor.ArtWorkflow & organisationStyle references
Tensor.Art Workflows and Models: Share, Reuse and Train
Tensor.Art lets you run and publish ComfyUI-style workflows in the browser, so a multi-step pipeline such as generate, upscale and face-fix can be saved once and reused by you or anyone else. Choose models by reading their cards and example data, train a LoRA from your own reference images when you need a repeatable style or subject, and record model versions and weights so results stay reproducible.
Beyond its one-click generator, Tensor.Art (tensor.art) hosts two things that working creatives tend to need as soon as they move past single images: shareable workflows, which are node graphs in the ComfyUI style that run on the site, and model training, which turns your own reference images into a LoRA. This guide covers how to pick models well, when a workflow beats the simple generator, and how to train and document a LoRA without wasting credits.
Choosing models by their cards
A model page on Tensor.Art gives you most of what you need to judge it before spending a credit.
- Base model family. SD 1.5, SDXL, Flux or another. Checkpoints and LoRAs only combine within a family, so filter by this first.
- Example images and their data. Open examples and read the prompt, negative prompt, sampler, steps, CFG, seed and LoRA weights. Creator-posted examples are the most trustworthy; community examples show the model’s range.
- Trigger words and recommended weight. Listed by the creator. Use them as written.
- Version history. Note which version the examples were made with. A newer version can look quite different.
- Usage permissions. Creators set what you may do with the output and the model itself. Read this before any paid work, and keep a copy.
For a style reference, look for a LoRA whose examples hold their look across different subjects. A LoRA that only looks right on portraits will not help you with a product shot. Test it on three dissimilar prompts at the same weight before adopting it.
When to use a workflow instead of the generator
The standard generator does one pass: prompt in, image out, with optional LoRAs and an upscale. A workflow chains steps. On Tensor.Art you can open workflows that other users have published, run them with your own inputs, and publish your own. Typical cases where a workflow earns its place:
- Generate, then upscale, then refine. A first pass at native resolution, an upscale, and a low-strength second pass that adds detail without changing the composition.
- Composition control. A ControlNet or similar node takes a sketch, pose or depth map so the layout matches a brief.
- Reference-driven style. An image-prompt or adapter node pulls the look from a reference picture instead of describing it in words.
- Face or detail fixing. A detector finds faces or hands and regenerates them at higher resolution.
- Batch variations. One graph that produces a consistent set from a list of prompts, useful for series work.
When you open a shared workflow, read the graph before running it. Identify which nodes take your input (usually a prompt, a reference image and a seed), and leave the rest alone on the first run. Then change one node at a time. Node names and the exact workflow editor layout may differ from desktop ComfyUI, but the logic is the same, so the ComfyUI beginner’s guide applies here too.
Training a LoRA from your own references
Tensor.Art offers on-site training, so you can make a LoRA without a local GPU. The quality of the dataset matters far more than the training settings.
- Collect 15 to 40 images of the subject or style. For a character, vary pose, expression, lighting and background while keeping the character identical. For a style, vary the subject while keeping the technique identical.
- Crop and clean. Remove text, watermarks and other people. Consistent framing helps less than consistent subject.
- Caption each image with what varies (pose, setting, clothing) and include a unique trigger word for what should stay constant. The trainer may auto-caption; review the captions anyway.
- Pick the base model you actually generate with. A LoRA trained on one family does not transfer to another.
- Leave the training defaults on your first run. Judge the result on the sample images the trainer produces at each stage, and pick the version where the subject is faithful but still follows new prompts.
Only train on images you own or have permission to use, and never on a living person without consent. For a text-only route to a repeatable character, see SeaArt consistent characters; the principles transfer.
Settings at a glance
| Setting | What it does | Start with |
|---|---|---|
| Base model filter | Shows only compatible models | Your checkpoint’s family |
| LoRA weight | Strength of a style or subject add-on | Creator’s value, else 0.7 |
| Workflow inputs | The nodes you change: prompt, reference, seed | Only these on a first run |
| Refine pass strength | How much a second pass may alter the image | 0.3 to 0.4 |
| Training dataset | Images the LoRA learns from | 15 to 40 clean, varied images |
| Trigger word | Unique token for your trained subject | A made-up word unlikely to exist in the model |
| Training epochs | How long the LoRA learns | Platform default, then pick the best sample |
Training and complex workflows cost more credits than a single image; check the platform’s current pricing page first.
Keeping it reproducible
A workflow with ten nodes has ten places for a setting to drift. Save a copy of any shared workflow you rely on, because the original can be edited or removed by its author. Record the checkpoint and LoRA versions, not just their names. When a trained LoRA works, note its base model, trigger word and best weight next to a few reference outputs. The test is always the same: could you rebuild this result next month from your notes alone?
Try it
Workflow: a published two-pass workflow (generate at native size, upscale 1.5x, refine at 0.35 strength)
Checkpoint: an SDXL checkpoint whose examples you like
LoRA: one style LoRA at 0.6 with its trigger word
Prompt: [trigger word], an architectural interior with a curved concrete staircase, morning light through a skylight, minimal furniture, wide-angle lens
Negative prompt: blurry, low quality, text, watermark, people
Seed: fixed
Run once as published. Then drop the refine strength to 0.25 and raise it to 0.5 on the same seed. You will see exactly how much freedom the second pass has, which is the one number that most changes how a workflow feels.