Astria

Where Astria sits

Astria is one of the older names in personalised image generation, and it has quietly narrowed its pitch since I first came across it. Where it once sold itself mainly as a "train a model on your face" service, the product today is pitched squarely at two workloads: ecommerce product and fashion photography, and professional headshots. The common thread is consistency — the same person, the same garment, the same product, rendered across dozens of scenes without a camera. The company is based in Tel Aviv and was founded in 2022 by Alon Burg and Itay Schiff, according to Dealroom and Tracxn profiles, and it has stayed relatively small and self-funded in a category crowded with well-funded rivals.

What separates Astria from the one-click avatar apps is that it is built around fine-tuning as a first-class, priced unit of work, and it exposes the whole thing as a REST API. That combination — a trainable subject model plus a documented API with callbacks and idempotency keys — is why a fair share of its users are not consumers at all but developers wiring personalised imagery into someone else's product.

What it actually does

The core loop is simple to describe. You upload roughly 10–20 images of a subject (a person, a product, a garment), Astria trains a fine-tuned model on them, and you then issue text prompts against that model to generate sets of images. On the web app this is packaged as an Ecommerce Studio — reusable photoshoot templates, on-model lookbooks, campaign scenes — and as headshot packs aimed at individuals and corporate teams. The API exposes the same primitives as two REST resources, tune (the fine-tuned model) and prompt (a generation request), nested as /tunes/:id/prompts.

Pricing is pay-as-you-go against an account balance, and it is unusually granular. At time of writing, the vendor's own pricing pages state $1.50 per fine-tune, $0.10 per prompt (a prompt returns 8 images), and $0.40 per 100 frames of video at 24fps. Beyond that, extras are metered individually: super-resolution or face correction at $0.0125, tiled upscale v2 at $0.07 per image, prompt masking at $0.01, and multi-person prompts billed as one prompt per person plus the base prompt. Training above 50 steps scales cost proportionally — 100 steps costs double — which is a pricing mechanic that rewards discipline with prompts. Minimum card top-up is $5, and credits sit in the account balance until spent.

Astria also now publishes a per-image menu for hosted foundation models — at time of writing it lists Gemini 2.5 (Nano-Banana) at $0.041 per image, Flux 2 Pro, Seedream 4.0 through 5.0 Lite, Recraft V4 and V4 Pro, and several in-house "Hyperrealism" and "Riverflow" models ranging from $0.008 to $0.275 per image depending on model and resolution. That menu tells you Astria has evolved from "bring your own subject" into a broader generation front end, but the fine-tuning workflow remains the differentiator.

The documentation covers the plumbing that production integrations need: bearer auth, callbacks (POST webhooks carrying the entity object), polling error semantics (422 validation, 429 rate limit, 500 retry with backoff, 504 with idempotency), and account-level idempotency keys so retried 504s do not double-create tunes or prompts. For non-API users, trained models can be exported as CompVis CKPT checkpoints or HuggingFace Diffusers archives.

My testing notes

Two concrete observations from working through the product and its docs:

First, the mock training mode is a genuinely useful engineering affordance. The API lets you create a Tune object with branch=fast for free; subsequent prompt calls return mock images at no charge. When I stepped through the docs, this stood out — it means a developer can build and test the entire integration (including callbacks, which work in test mode) before spending a cent. Very few fine-tuning APIs bother with this; most charge you to find out your webhook payload shape is wrong.

Second, the storage policy forces a real workflow decision. Trained models, training images, and generated images are deleted 30 days after training completes unless you pay $0.50 per model per month for extended storage or set auto_extend on the tune. In a headshot workflow where a client comes back a quarter later, that means either keeping a $0.50/month tab open per subject or retraining at $1.50. The pricing is small, but the retention design shapes how you build on top of it.

Third, the billing mechanics reward batching. A prompt returns 8 images for $0.10, so per-image cost collapses to about $0.0125 for fine-tuned output before extras — but every uplift (face correction, tiled upscale, prompt masking) is metered on top, and the docs are explicit that steps above 50 multiply cost. In my read of the pricing model, a naive integration that always requests face correction and 100 steps can spend 10x what a disciplined one does on identical output volume.

One more thing I checked: the fine-tuning handbook and gallery pages show what good output looks like — fashion lookbooks, sports apparel, eyewear campaigns — and the quality bar is credible for catalog and headshot work. The homepage examples are vendor-selected, so I treat them as a ceiling, not a median, but the API's pricing transparency (each addon enumerated in cents) is the kind of detail that usually correlates with a product built for repeat commercial use.

What it does not do

Astria is not a design tool — there is no canvas, no layout, no brand kit. It generates; you compose elsewhere. It does not support nude content, which is a hard policy limit. It explicitly declines to support using exported checkpoints with external tooling ("Astria does not provide support for using external tools with the checkpoints"), so if your plan is to train cheaply and run the model yourself on ComfyUI, you are on your own. There is no free tier in the normal sense — the mock branch is for testing, and real usage starts at a $5 top-up. Occasional demand surges can delay prompt processing; the vendor itself tells users to watch the ETA page on /tunes, which tells me queue times are a known friction point. And while the foundation-model menu broadens the surface, Astria cannot match the raw model breadth of aggregator workspaces like getimg.ai.

Positioning against the avatar apps

I want to be clear about the competitive frame, because buyers compare Astria with one-click headshot mills and with general image workspaces, and it is neither. The headshot mills sell a fixed pack of 40 to 100 images for a one-time fee and give you no control over the pipeline. General workspaces give you every model but no durable subject identity. Astria sits between them: you pay for training and generation separately, keep the tune alive as long as you pay for storage, and can regenerate a subject's lookbook six months later with a new prompt set. For an ecommerce operator refreshing seasonal campaigns on the same models and garments, that continuity — not any single image — is the product.

Verdict

Astria earns its place for a specific buyer: someone who needs the same face or product rendered consistently at volume, and who is comfortable in an API-first, metered-everything environment. The fine-tuning workflow is mature, the pricing is legible down to the cent, and the developer ergonomics (mock mode, idempotency, callbacks, checkpoint export) are better than most of the avatar-app field. Against that, there is no free tier, model retention costs money after 30 days, and the product asks for prompt-engineering discipline to stay cheap. I rate it 3.8 — strong core depth in fine-tuned consistency and a well-drawn API (the 40% functionality pillar), fair pay-as-you-go value (20%), a usable if sparse web app (20%), modest ecosystem beyond the docs (10%), and generally positive reliability signals with acknowledged queue delays (10%).

Sources

• https://www.astria.ai/
• https://astria.ai/pricing/sd?ws=null
• https://www.astria.ai/pricing
• https://docs.astria.ai/docs/api/overview
• https://astria.ai/api/docs
• https://tekpon.com/software/astria-ai/reviews
• https://dealroom.co/companies/astria_ai
• https://tracxn.com/d/companies/astria/__OpJEbjUDi7N4V0iTzGQWqZwBMBEWTuCGgqlfkeI4bqw
• https://effecthub.net/ai/astria
• https://startupim.com/company/astria
• https://docs.astria.ai/docs/use-cases/ai-photoshoot/