protoBanana

💬

Conversational by default

One gateway alias drives multi-turn editing — "draw a cat" → "now make it blue" → "remove the background" — all auto-routed per turn.

🔌

Drop-in OpenAI shape

Hits /v1/chat/completions, /v1/images/generations, /v1/images/edits exactly like DALL-E or Nano-Banana 2. Open WebUI, protoCLI, raw curl — same call, every client.

🛠

Composable workflows

Each operation (gen, edit, multi-ref, sticker, region edit, inpaint, outpaint) is one ComfyUI workflow JSON + one Python route. Adding capabilities is mechanical.

🧠

Backed by SOTA OSS models

Qwen-Image-2512 + Qwen-Image-Edit-2511 (gen+edit, multi-ref) + BiRefNet/RMBG-2.0 (sticker) + Florence-2 + SAM 2.1 (region edit, Phase 4) + LanPaint (inpaint, Phase 5).

🏠

All your data, all local

For organizations that can't or won't send data to a third party. Your gateway, your ComfyUI, your weights. We never see a pixel.

📈

Honest about gaps

5-15pp behind frontier on hardest cases. 3-reference cap (Qwen ceiling) vs Nano-Banana 2's 14. Where it's good, it's competitive; where it's not, we say so.

What you get

# In your chat client (Open WebUI, protoCLI, or raw OpenAI SDK):

  user: a watercolor of a cat in a hat, portrait
  [image: cat in hat, 832×1216]

  user: now make it blue
  [edited image]

  user: remove the background
  [transparent png]

  user: change just the hat to red
  [masked region edit — Phase 4]

One model alias (protolabs/qwen-image-chat) handles all of it. The provider walks message history, classifies the operation per turn, dispatches to the right ComfyUI workflow.

When to use it

Use protoBanana when	Use Nano-Banana 2 / GPT-Image-2 when
Data sovereignty / compliance / IP sensitivity	You don't care where the data goes
You want fixed cost (electricity) at scale	You're under metered-API-call budgets
You need to extend with custom workflows	Frontier-quality output is non-negotiable
You already run a LiteLLM gateway	You don't have GPU infrastructure

Use protoBanana when

Use Nano-Banana 2 / GPT-Image-2 when

Data sovereignty / compliance / IP sensitivity

You don't care where the data goes

You want fixed cost (electricity) at scale

You're under metered-API-call budgets

You need to extend with custom workflows

Frontier-quality output is non-negotiable

You already run a LiteLLM gateway

You don't have GPU infrastructure

For most teams: both. Use the closed APIs for one-off best-quality work, route bulk + sensitive workflows through protoBanana.

protoBananaChat-native image gen + edit. Open-source. Local.

Conversational by default

Drop-in OpenAI shape

Composable workflows

Backed by SOTA OSS models

All your data, all local

Honest about gaps

What it is, in one sentence

What you get

When to use it

Where to go next

protoBananaChat-native image gen + edit. Open-source. Local.

Conversational by default

Drop-in OpenAI shape

Composable workflows

Backed by SOTA OSS models

All your data, all local

Honest about gaps

What it is, in one sentence ​

What you get ​

When to use it ​

Where to go next ​

What it is, in one sentence

What you get

When to use it

Where to go next