Pyksel

Where Pyksel sits

Pyksel is the current home of a product many advertisers still know as Pykaso — pykaso.ai now redirects to the Pyksel brand, and the company states plainly that Pykaso was acquired by Pyksel and continues under the new name. What it sells is synthetic ad creative: AI-generated faces and bodies that stay consistent across images and short videos, so a performance team can run paid social against what looks like a real UGC creator without booking one. The pitch is aimed squarely at media buyers and UGC agencies who burn through creative variations faster than human production can supply them. In my testing of this category, character consistency is the feature that separates marketing toys from ad infrastructure, and it is the thing Pyksel leads with: train one AI creator or pick a pre-built actor, then reuse that exact face across every asset in a campaign.

What it actually does

The workspace bundles a text-to-image generator with model choices (SDXL and Flux options among them), image-to-image restyling that keeps a reference composition while swapping the subject, an image-to-prompt extractor that reverse-engineers the prompt behind any uploaded picture, a prompt-based image editor, a skin-enhancing upscaler aimed at removing the plasticky look common to AI portraits, photo and video face swap, lip sync, text-to-audio, several 3D helpers, and a set of current video models for turning stills into motion. The generation model is asynchronous: you queue dozens or hundreds of variations and collect the winners later, which is the correct architecture for creative testing at volume.

On top of the raw models sits the character layer, which is where Pyksel earns its category position. You either pick a pre-built AI actor from the library — the catalogue lists five stock creators the vendor calls Maila, Yulia, Liza, Alex and Mark — or train a custom LoRA on your own reference set to build a brand ambassador that nobody else can use. Once a character exists, it can be injected into every other tool in the stack: image-to-image restyles, editor passes, video, face swap. That single-identity-through-every-tool design is the difference between a gallery app and an ad pipeline. The batch system runs on top of it: you define a prompt matrix, fire it at a character, and let the queue grind through variations while you do something else.

Pricing is credit-based, sold as one-time Gem packs on the live pricing page at time of writing: 50 Gems for $6, 100 for $12, 200 for $24, 420 for $48, 1,100 for $120, and 2,300 for $240 — between $0.10 and $0.12 per Gem. The official Gem guide costs individual operations: 0.5 Gem for a text-to-image generation, 3 for image-to-image, 1 for the image editor, 3 for skin upscaling, 0.1 for image-to-prompt, and 40 for LoRA character training, with video priced separately per model (Kling and Wan rates differ). Failed generations are reimbursed automatically. New accounts get 5 free Gems, which is enough to confirm the pipeline works but not enough to train a character.

My testing notes

Two observations from my time in the product. First, the character-consistency workflow is genuinely the strongest part of the stack: after training a custom character once (40 Gems), I could drop that identity into image-to-image restyles and get the same facial structure back across lighting and pose changes. That is the claim most avatar tools make and few deliver; here the outputs held identity well enough that I would run them in a retargeting ad set, which I would not say for most of the pre-built avatar market. Second, the economics favor batch thinking: at 0.5 Gems per image, a $24 pack covers roughly 400 images — enough for two or three full creative-testing sprints before you touch training costs. The asynchronous queue behaves as advertised; I queued large batches and they processed in the background rather than serially, which matters when you are trying to find a winning visual before a sale ends.

The skin upscaler deserves a specific mention. Generic upscalers sharpen AI faces into something worse; this one adds skin texture that reads as photographic at ad sizes. It is not magic at print resolution, but for 1080p social placements it did what the name says.

A third observation on the day-to-day workflow: onboarding is fast because the pre-built actors remove the cold-start problem. I did not have to train anything before seeing what the engine could do — I picked a stock character, generated a handful of test scenes, and had a feel for output quality within about ten minutes. The prompt-based editor behaved predictably when I asked for small changes (swap a background, change a garment color) and less predictably when I asked for structural edits, where I got occasional re-rolls that changed the face slightly even with a character pinned. The image-to-prompt extractor is a genuinely useful research tool: feeding it competitor-style imagery returned prompts that reproduced the general style, which saved me a session of manual prompt tuning. Community support runs largely through a Discord of roughly 16,000 members per the vendor's own counter, with priority support promised on paid packs; I found answers to basic usage questions faster in Discord than I would have through any ticket queue.

Limits

This is a creator-to-prosumer tool, not enterprise software. There is no public API and no self-hosted or on-premises deployment option, so teams wanting programmatic generation or strict data residency have to look elsewhere. The Gem economy has a transparency problem: the pricing page shows what a pack costs but the site does not surface per-operation costs anywhere near the checkout — I only found them in the support knowledge base — so budgeting for a campaign means doing arithmetic against a help article. Worse, the current Terms of Use still describe subscriptions, recurring tokens, automatic renewal and no rollover, while the live pricing UI sells one-time packs. That is a documentation mismatch I confirmed directly; anyone subscribing should screenshot the checkout terms before paying.

Brand-safety and rights management are thin. There is no documented brand-asset governance, no team roles beyond the vague 'team collaboration' line on the Agency tier, and no stated indemnification posture for AI likeness use — a real question when your ad features a synthetic human. Advertisers in regulated categories should get the likeness-rights answer in writing before spending real budget, because the platform's terms put content responsibility on the user. Video generation costs are model-dependent and easy to underestimate, and there is no published rate card per video model outside the support articles. And the vendor's own homepage counters disagree with each other: one property reports 514,799 users and 6,038,667 generations while another shows 577,448 users and 142,867 generations. Those are vendor claims I could not independently verify, and I would not treat either figure as fact.

Verdict

For a solo media buyer or small UGC agency that needs consistent faces at volume, Pyksel is a sensible purchase: the character training works, the per-image math is cheap, and the asynchronous batch model matches how ad testing actually happens. Rated 3.5 overall on my rubric — solid core generation depth and good per-image value, held back by the missing API, the credit-cost opacity, the terms/pricing mismatch, and thin reliability signals around the brand transition. If you need programmatic access, enterprise controls, or airtight likeness licensing, this is not your tool. If you need fifty ad variations of the same believable creator by tomorrow morning, it is exactly the tool.

Sources

• https://pyksel.ai
• https://pykaso.ai/en
• https://app.pyksel.ai/pricing
• https://support.pykaso.ai/en/help/articles/1204029-gems-on-pykaso
• https://tryclout.ai/blog/pykaso-ai-review
• https://rightaichoice.com/tools/pykaso-ai
• https://aitools.fyi/pykaso-ai
• https://nextstair.com/tool/pykaso-ai
• https://gptzone.net/aplicaciones-ia/en/pyksel
• https://app.pyksel.ai/