Image editing API · served now
Instruction-based editing through the OpenAI-compatible /v1/images/edits endpoint, served from hardware AxForge owns in Sweden. Send an image and a sentence — swap the jacket, replace the sky, remove the crowd — and get the edit back in seconds. Zero prompt retention.
Specifications
| Model | FLUX.2 Klein 4B — instruction-based editing with multi-reference support |
|---|---|
| API model name | flux2-klein-4b |
| Endpoint | POST /v1/images/edits |
| Input | Source image(s) (base64 or data URL) + a plain-language instruction |
| Typical edits | Clothing and color swaps, sky and background replacement, object removal or addition, product-shot restyling |
| Text in images | A generation job — regenerate with ERNIE Image Turbo, built for dense, legible in-image text |
| License | Apache 2.0 (open weights, 4B variant) |
| Quantization | fp8 |
| Hardware | NVIDIA DGX Spark (GB10, 128 GB unified memory) — owned and operated by AxForge |
| Region | eu-se-1 · Stockholm, Sweden — pinned with x-axforge-region, echoed on the response |
| Pricing | €0.02 / image (launch pricing) |
Picked in a same-source, same-instruction bake-off on our own hardware: near-parity with the previous edit model on visual edits, at three times the speed and a third of the memory. Image generation runs on ERNIE Image Turbo at €0.02 per image.
Performance
| Metric | Value | Condition |
|---|---|---|
| Edit time | ~8 s | per instruction edit, 4-step distilled, fp8 serving |
Measured on our production DGX Spark node, single-stream, 2026-08. We publish only numbers we measured ourselves.
Quickstart
$ export OPENAI_BASE_URL=https://api.axforge.ai/v1 $ curl -sS "$OPENAI_BASE_URL/images/edits" \ -H "Authorization: Bearer $AXFORGE_API_KEY" \ -H "x-axforge-region: eu-se-1" \ -H "Content-Type: application/json" \ -d '{"prompt":"replace the grey sky with a clear summer evening", "image":"'"$(base64 -w0 photo.png)"'"}'
Data & privacy
Images and prompts are processed in memory in Sweden — not written to disk, not logged, not retained, and never used to train anything. We keep only request metadata (token counts, timestamps, status) for billing and operations. The full policy is at axforge.ai/privacy.
FAQ
The classic photoshop-style requests, from a plain-language instruction: swap clothes or colors on a person, replace a sky or background, remove or add objects, restyle a product shot. Send the source image plus one sentence. Multiple reference images are supported.
Text belongs to the generation model: regenerate with ERNIE Image Turbo, which is built for dense, legible in-image text. Keep Klein for the visual work — objects, backgrounds, clothing, style.
We measured roughly 8 seconds per edit (4-step distilled model) on our production DGX Spark node, single-stream, 2026-08. We publish measured numbers only.
€0.02 per image (launch pricing). No subscription, no minimum. For committed volume, talk to an engineer.
The 4B model we serve is Apache 2.0 — fully open for commercial use. The larger 9B variant carries a non-commercial license; we do not use it.
No. Images and prompts are processed in memory in Sweden, never persisted or used for training — see the privacy policy. This makes the API a fit for GDPR-sensitive workloads.