Alibaba's 20B open-source editing model with dual-path architecture — natural language edits, multi-image composition, and bilingual text rendering at $0.03/megapixel.
Explore all variants in the Qwen Image Edit family.
Latest 20B editing model with dual-path architecture
Customizable variant with LoRA support
Multi-angle view generation
Enhanced editing capabilities
Every model on Runflow is scored per vertical — not generic leaderboard numbers. Scores come from domain-expert evaluation on real production prompts.
| Use Case | Assessment | Notes |
|---|---|---|
| Portrait | Pending | Portrait editing quality |
| Fashion | Pending | Fashion editing quality |
| Product | Pending | Product editing quality |
| Creative | Pending | Creative editing quality |
Full parameter reference for the qwen-image-edit-2511 model endpoint.
| Parameter | Type | Description |
|---|---|---|
| prompt | string | Natural language instruction describing the edit. Examples: 'change the background to a beach', 'make the lighting warmer'. |
| image_urls | array[string] | URLs of the source images to edit. Accepts 1–3 images for reference-based or multi-image editing. |
| negative_prompt | string | Describe what to avoid in the edited output — artifacts, unwanted styles, or quality issues. |
| acceleration | string | Speed/quality tradeoff. 'none' for maximum quality, 'regular' for balanced, 'high' for fastest editing. |
| image_size | object | string | Output dimensions. Defaults to input image dimensions if not specified. |
| num_images | integer | Number of edited variations to generate per request. Range: 1–4. |
| num_inference_steps | integer | Number of denoising steps. 20 for fast edits, 28 for balanced, 35–50 for max fidelity. Default: 28. |
| guidance_scale | number | Controls how closely the model follows the edit instruction. 3–5 for subtle, 6–10 for dramatic. Default: 4.5. |
| seed | integer | Reproducibility seed. Same seed + same inputs = identical output. |
| output_format | string | Output image format. Options: "jpeg", "png", "webp". Default: "png". |
| enable_safety_checker | boolean | If set to true, the safety checker will be enabled. |
| sync_mode | boolean | If true, returns the image as a data URI. |
Flux AI models are available on many platforms. Here's what makes the Runflow image generation API different.
Unique two-stream design: Qwen2.5-VL captures semantic meaning (what to change), while a VAE encoder preserves appearance fidelity (what to keep).
Accepts 1–3 input images for person+person group photos, person+product placement, person+scene integration, and style transfer between images.
Add, remove, or modify text within images in both English and Chinese with natural typography — spacing, alignment, and style consistency maintained.
Maintains facial identity and character consistency across edits — critical for portrait retouching, pose changes, and multi-person group photo fusion.
Apache 2.0 licensed with full weights on HuggingFace. The largest open-source image editing model, self-hostable and customizable via LoRA fine-tuning.
Same API, different strengths. Switch models with one parameter change.
Black Forest Labs
FLUX.1 [dev] is a 12B-parameter text-to-image model by Black Forest Labs. Open-w…
ByteDance
Seedream 4.5 by ByteDance — unified generation and editing in one model. Strong…
ByteDance
Seedream 4 by ByteDance — unified generation and editing model with photorealism…
Nano Banana 2 by Google — fast image generation across 4 resolution tiers (0.5K–…
OpenAI
GPT-Image 1.5 by OpenAI — high-fidelity images with strong prompt adherence. Thr…
Nano Banana Pro by Google — premium image generation with sharp detail and natur…
Get API access in minutes. No GPU setup, no infrastructure to manage. Pair with Sentinel to control the quality of every AI-generated image you ship.