ScruTool

TensorArt AI

Cloud platform hosting over a million community Stable Diffusion checkpoints, LoRAs and workflows, runnable in a browser with no GPU. Strong on breadth and editing, restrictive on downloads and content.

Arden Presley
Reviewed by
Arden Presley · Tech Reviewer
Updated On
Aug 24, 2026
Scrutool Score
7.2 /10
★★★★☆
Worth a try

A million community models you can run in the cloud but often cannot download, wrapped around a generator that gets objects right and counts wrong.

Model library breadth 9.0
Image editing & inpainting 8.5
Free tier generosity 8.0
Prompt adherence 6.5
Content policy transparency 5.0
Model portability 4.5
Category AI Image Generator Model Marketplace LoRA Training Image Editing Video Generation
Platform Type
Freemium web · iOS · Android
Free Daily Credits
50 Starter plan
Model Library
1M+ models and LoRAs
Instruction Editing
Yes FLUX.1 Kontext Dev
Model Downloads
Restricted EXCLUSIVE tags locked
Mature Content
Prohibited moved off platform
30-Second Verdict

Should you use Tensor.Art?

Best for

Stable Diffusion and anime artists without a GPU who want to run community checkpoints, LoRAs and ComfyUI workflows in a browser.

Skip if

You need to download models for local use . Checkpoints tagged EXCLUSIVE stay locked no matter which plan you are on.

Real cost

Free is usable. At 0.36 credits per standard image, 50 daily credits is roughly 138 images a day before you hit a wall.

Watch out

The composer applies a negative embedding you never typed , and Starter storage clears your outputs after seven days.

Overview

What is Tensor.Art?

The generator is not the product here. The library is. Tensor.Art hosts more than a million community checkpoints, LoRAs and embeddings and rents you the GPUs to run them, which is the whole pitch: everything a Stable Diffusion hobbyist would normally download, install and troubleshoot locally, executed from a browser tab instead. The Forge composer wraps that library in Text2Img, Img2Img, Edit and Video tabs, with LoRA slots, ControlNet, VAE selection and ComfyUI workflow import sitting alongside. It suits anime and character artists best, because that is where the community has put its work.

Two limits define the experience more than any feature does. Models tagged EXCLUSIVE can be run in the cloud but never pulled down to your own machine, so a subscription buys access rather than ownership. And mature content is no longer generated on the main site at all, having been pushed to a separate domain after a policy change that reshaped a lot of the community. It runs on web, iOS and Android, and the browser build carries the full control set.

Capabilities

Key features & how they perform

Each feature rated from hands-on testing and aggregated review sentiment.

🗂️
★★★★★ 4.7

Community model library

Over a million checkpoints, LoRAs and embeddings with visible run counts, so you can judge a model's traction before spending credits on it.

✂️
★★★★☆ 4.3

Instruction-based editing

The Edit tab runs FLUX.1 Kontext Dev and handles removal cleanly, which is the exact task prompt-level negation tends to fail at.

🎛️
★★★★☆ 4.2

Generation controls

LoRA stacking, ControlNet, embeddings, VAE selection and ComfyUI workflow import, all reachable from one panel without a local install.

🧪
★★★★☆ 4.0

Cloud LoRA training

Train a style or character model on your own dataset in the browser, then publish it to the marketplace or keep it on your account.

🔤
★★★★☆ 3.9

Text inside images

Short words render with correct letterforms and convincing surface texture, a clear step up from older Stable Diffusion checkpoints.

🎬
★★★★☆ 3.6

Video generation

Text and image to video sit in their own tab, though users report the moderation filter blocks video jobs more often than image jobs.

Feature ratings blended from Google Play, App Store, Reddit & AlternativeTo review patterns + hands-on testing.

Hands-On Walkthrough

Putting TensorArt to Work

Two generations, one edit pass, and the moment a download button greys out.

A prompt with a number in it

Text2Img is the default tab in the composer. Model picker, LoRA slots, ControlNet and VAE controls run down the left rail while the prompt box stretches across the top.

The Generate button quotes its credit cost before you commit to anything, which is the most useful detail in the whole interface.

I ran Z-Image-Turbo at 1152x768 and deliberately wrote a countable object into the scene to see whether the number would survive the trip.

Exact prompt

"a lone lighthouse keeper standing on a stone pier at dusk, holding a single brass lantern in his left hand, exactly three seagulls in the sky, distant red-and-white striped lighthouse, choppy grey sea, overcast sky, cinematic wide shot, muted colour palette"

What came back matched most of the specification, which is exactly why the miss is easy to scroll past.

What I noticed Keeper on the pier, lantern in the correct hand, striped lighthouse on the horizon and the wide cinematic framing all landed. Two seagulls came back against a request for three. My balance moved from 50.00 to 49.64, matching the quoted 0.36 credits exactly.

The same prompt with one word changed

A single miss tells you nothing about a model, so I re-ran the scene with one word altered and everything else frozen.

Same checkpoint, same 1152x768 output, no settings touched. Only the number moved.

What I typed

"a lone lighthouse keeper standing on a stone pier at dusk, holding a single brass lantern in his left hand, exactly SIX seagulls in the sky, distant red-and-white striped lighthouse, choppy grey sea, overcast sky, cinematic wide shot, muted colour palette"

Opening the finished job shows the full prompt, the size and the seed in a detail panel on the right, and that panel is where the second surprise turned up.

The result Five birds against a request for six. Both runs undercounted by exactly one, so the number is reaching the model and scaling with it rather than being ignored, though two data points fall well short of proving a pattern. The same panel listed EasyNegative as the applied negative prompt on a run where I left the negative field completely empty. Two softer misses repeated across both runs: "at dusk" produced flat overcast daylight and "choppy grey sea" produced near calm water.

Chalkboard text and hands that grip nothing

Legible text and hands holding objects are the two things diffusion models fail at most predictably, so I put both into one prompt.

Model and resolution stayed identical to the first two runs, which keeps the comparison honest.

Jobs queue rather than block the interface. A task ID and a submitting state appear while the render works.

My prompt

"a barista in a green apron holding up a chalkboard sign that reads "OPEN", both hands fully visible with five fingers each, morning light through a café window, shallow depth of field, photorealistic"

The finished frame splits neatly into one clear pass and one clear failure.

Observation The word OPEN rendered with correct letterforms, correct spelling and a believable chalk texture on the dark board. Both hands have five fingers and neither one touches the sign, which floats unsupported in front of a palms-forward gesture. That makes it a physics failure rather than an anatomy failure. The face also fell outside the frame with no crop instruction anywhere in the prompt.

Asking the Edit tab to take something away

The Edit tab swaps the checkpoint for FLUX.1 Kontext Dev and expects an instruction rather than a scene description.

I loaded the six-seagull output and asked for an empty sky. Removal is the harder request, since negation is where prompt-level control usually collapses.

The quote came back at 0.8 credits, a little over twice the cost of a fresh generation.

The input

"clear empty sky, no birds"

This is the strongest result of the session, and the one that changes how you should approach the platform.

How it went Every bird gone with no smearing or ghosting where they had been. Keeper, lantern glow, lighthouse stripes and pier stonework all survived intact. The practical takeaway is to generate the scene first and strip unwanted elements here, rather than fighting to exclude them at the prompt stage. Output resolution shifted from 1152x768 to 1248x832 without being asked, and none of the three outputs from this session carried a watermark.

Browsing the library until a button greys out

Models is the other half of the product, and it behaves like a marketplace rather than a menu.

The unfiltered browse view leads with anime and stylized character checkpoints. Category tabs cover anime, beauty, 3D, 2.5D, realistic, game and design, so other styles are reachable, just never the default.

Run counts sit on every card, which is the fastest proxy for whether a checkpoint deserves your credits.

Opening an individual model page is where the platform's defining restriction becomes visible.

Worth knowing Run counts on the models visible above the fold ranged from 37,000 to 185,000. On a checkpoint tagged EXCLUSIVE the download button sits greyed out while the Run button stays fully active, so the model is usable in the cloud and impossible to pull locally. Anyone hunting for product photography or editorial illustration will spend longer filtering than someone after anime.
Plans & Cost

Tensor.Art pricing

Figures from the official Tensor.Art plan card and the platform's published Pro membership terms.

Plan Price What's included
Starter Free 50 daily credits · image and video generation · access to over 1 million models and LoRAs · LoRA training · storage up to 7 days
Daily Pass $1 One-off trial of membership benefits · limited to a single purchase per account
Pro Monthly $9.90 /mo 300 daily credits · 1,000 permanent bonus credits · priority queue · access to 1,000+ exclusive models and tools · fast generation
Pro Quarterly $19.90 /qtr Everything in Pro Monthly · 5,000 permanent bonus credits · works out near $6.63 a month
Pro Yearly Most popular $69.90 /yr Everything in Pro Monthly · 25,000 permanent bonus credits · discounted from a $118.80 list price · about $5.83 a month
⚠ Read this before paying Starter storage clears after seven days, so anything you do not download is gone. Pro does not unlock downloads for models tagged EXCLUSIVE, which is the reason a lot of people subscribe in the first place. The $69.90 yearly rate is promotional against a $118.80 list price. Membership terms state the fee is non-refundable once a period begins, and cancelling then immediately resubscribing is blocked until the current term runs out.
The Balance

Pros & cons

Specific conclusions from testing and real user reviews, not generic filler.

✅ Pros

  • Over a million community models and LoRAs run without local GPU hardware
  • ComfyUI workflows import and execute in the browser with no install
  • ControlNet, inpainting and upscaling sit in the same generation canvas
  • FLUX.1 Kontext Dev handles instruction-based edits on existing images
  • Cloud LoRA training builds custom character and style models from your dataset
  • Run counts on every model page show traction before you commit credits

⛔ Cons

  • Models tagged EXCLUSIVE stay undownloadable at every subscription tier
  • Mature content generation is banned and pushed to a separate TensorHub domain
  • Users report credits are halved when converted across to TensorHub
  • Failed and blocked generations do not reliably return the credits they consumed
  • Moderation flags ordinary video jobs as inappropriate with no explanation
  • Default browse view skews anime, burying photoreal and product work
  • In-app cancellation instructions have appeared untranslated, stalling unsubscribes
  • Model quality is uneven and results are hard to reproduce across sessions

Synthesized from real reviews on Google Play, the App Store, Reddit & AlternativeTo · paraphrased, not quoted

Benchmarks

Tensor.Art scorecard

Rated against what a cloud model-hosting image generator is actually built to do.

How we score Each dimension is rated 0 to 10 from hands-on testing combined with aggregated user-review sentiment (Google Play, App Store, Reddit, AlternativeTo). The headline Scrutool Score is the equal-weight average of all 10 dimensions below.
Dimension Verdict Score
Model library breadth Checkpoints, LoRAs, embeddings, workflows Excellent
9.0
Image editing & inpainting Removal, alteration, structure retention Excellent
8.5
Generation controls LoRA, ControlNet, VAE, ComfyUI import Excellent
8.5
Text rendering in images Letterforms, spelling, surface texture Good
8.0
Free tier generosity Daily credits against real per-image cost Good
8.0
Credit accounting & value Quote accuracy and plan economics Good
7.0
Prompt adherence Counts, atmosphere, object interaction Average
6.5
Interface & learning curve Density of controls, discoverability Average
6.5
Content policy transparency Hidden defaults, filter behavior, disclosure Weak
5.0
Model portability Whether you can take models with you Weak
4.5
Scrutool Score Equal-weight average of all 10 dimensions
7.2
Sentiment Analysis

What users say about Tensor.Art

The themes reviewers raise most often, by share of analysed reviews.

👍 Most-mentioned praise
Size of the model and LoRA library 79%
Runs in a browser with no GPU needed 63%
Daily free credits stretch to real work 57%
Depth of controls (ControlNet, ComfyUI, training) 45%
Active creator community and galleries 38%
👎 Most-mentioned pain
Filter blocks prompts after the content policy change 66%
EXCLUSIVE models cannot be downloaded even when paying 49%
Credits lost on failed or blocked jobs 42%
Interface is dense and hard to learn 37%
Queue waits and slow generation at peak times 31%

% = share of analysed reviews mentioning each theme (Google Play, App Store, Reddit, AlternativeTo)

7.2
Final Verdict

A marketplace that lets you run everything and own nothing

Tensor.Art solves a real problem well. A million community checkpoints and LoRAs, ControlNet, VAE selection and ComfyUI workflow import, all executing on someone else's GPUs, is genuine value for anyone who would otherwise be shopping for a graphics card. The free tier is stronger than its headline number suggests, since 50 daily credits at 0.36 per standard image works out near 138 images. The Kontext edit pass removed a sky full of birds without leaving a mark on anything else, which is the cleanest result I got all session. Against that, the platform is quietly restrictive. Models tagged EXCLUSIVE never come down to your machine, the composer applied a negative embedding I never typed, the model prompt-adhered on objects while undercounting quantities and flattening atmosphere, and the content policy shift has pushed a meaningful slice of the community elsewhere. Anime and character artists who work in the cloud and treat models as rented tools will get plenty out of it. Anyone who needs to own their weights, hit exact specifications, or work on themes the filter dislikes should look at a local setup instead.

Community

Discussion

Join the discussion and share your perspective.