One tool.
Every kind of
generation.
Text to image, image to video, reference to video, face and character swap, lip-sync, upscale — one balance, one interface, no separate tools to learn. Bring your own LoRAs, and render on data-centre GPUs instead of your own.
Browse every model and its price before you sign up.
text → image
text → image
text → imageHow it works
No install. Pick a model, drop in a prompt or a source image, and you're looking at results in a few minutes — or seconds if a GPU is already warm.
Give it something
A prompt, a photo, a short clip, or an audio track — whatever the mode calls for. Your own LoRAs load per model.
Pick the mode
Text to image, image to video, reference to video, swap, lip-sync, upscale — same interface, same balance.
Keep what you make
Full resolution, no watermark, yours to use. Unused Takes never expire. Failed jobs refund themselves.
Want a shot that happens in beats - a door opens, then a cat walks in, then it meows? Script it by time, with sound →
Under the hood
Several open models, picked per mode, so you're not stuck with one engine for everything. Each runs on data-centre hardware rented by the hour — which is also why it's priced the way it is.
MiniMax H3
Video with synchronised audio, in a single pass.
Wan 2.2
Video. Animate a still, or put your character into a performance.
Krea2 Turbo
Images. Our default — fast, and the one model with a style LoRA.
Flux 2 Klein
Images. 4-step draft speed.
Qwen-Image
Images. Generate from text, or edit an image you already have.
InfiniteTalk
Lip-sync. Make a photo or a clip talk, driven by audio.
FaceFusion
Face swap. Graft one face onto an image or video.
SeedVR2
Upscale. Restore and enlarge video to 1080p, 2K or 4K.
LoRAs load per model — pick your base first, then the LoRA trained for it. Civitai and Hugging Face links, or a .safetensors upload.
Data-centre GPUs, not your laptop
Every render runs on rented data-centre cards — the L40S for images and Wan video, RTX PRO 6000 Blackwell class for MiniMax H3. Nothing to install, no VRAM to budget, no fan noise, and your own machine stays free while it renders.
~15s
an image at 720p
3–4 min
a 4-second clip, including frame interpolation
80 GB+
of VRAM on every card that runs H3 — a model that does not fit on any consumer GPU
Measured render times on the cards the fleet rents, once a GPU is warm. A cold start adds a few minutes while one boots.
BYOL — bring your own LoRAs
Use it like your own ComfyUI box. Paste any LoRA link from Civitai or Hugging Face — or upload a .safetensors — and it loads onto the model you picked. No approved list, no waiting for us to add it, no curated styles you are stuck with.
A style, a character, a subject — anything Civitai or Hugging Face hosts for that base model. Stack several, set each one's strength.
Resolution, duration, aspect, seed and the step-level options each model exposes — the knobs you'd reach for locally, not one fixed preset.
Everything runs through ComfyUI graphs on our fleet, so new generation types ship when we build them — not when a vendor's API catches up.
$5 = 1,200 Takes.
An image is 1. A 4-second clip is 16.
Takes are prepaid credit. Every generation shows its price before you run it; that is what you pay. Bigger top-ups buy more per dollar: 120/$ from $5, 140/$ from $20, 160/$ from $50.
Start with $5USD shown at the best rate (160 Takes/$). API use is priced above the web rate.
Made with $5
Questions
- What is a Take?
- The prepaid unit every generation is priced in. $5 buys 600 Takes (160 per dollar at the best tier). An image is 1 Take; a 4-second 480p video clip is 16 Takes. Every generation shows its price before you run it, and Takes never expire.
- Do I need a subscription?
- No. Ersolin is prepaid. Pro ($29/month for 4,500 Takes and 5 concurrent generations) is optional.
- Can I use my own LoRAs?
- Yes. Paste a Civitai or Hugging Face link, or any direct .safetensors URL, on models with LoRA support. They load per generation; there is no curated list to wait for.
- What happens if a generation fails?
- The Takes are refunded automatically and you are told by email. Cancelling a queued generation refunds it; cancelling one that is already rendering does not.
- Is there an API?
- Yes. Create a key on /developer (it starts with ers_live_) and call the /v1 endpoints with a Bearer token. API generations cost 1.5x the web price, rounded up.
- Who runs Ersolin?
- Ersolin is a trading name of Racoon2e. Contact hello@ersolin.com.