2026-03-03 · Lab Notes ◆ Experimental
Image Model API Baseline
9 models. 3 providers. 2 prompts. One clear recommendation.
Models 9
Providers 3
Pass Rate 72%
Driver Nano Banana 2
Speed
The BFL Problem
Fireworks Flux
100% pass — IP works
⇆
BFL Flux
0% pass — IP blocked
Same model family. Different provider. Opposite reliability. BFL content filters block IP character prompts (Ninja Turtle, Cinderella). Fireworks passes them all.
Reliability Matrix
| Model | Character IP | Environment | Rate |
|---|---|---|---|
| nano-banana | PASS 6.5s | PASS 4.9s | 100% |
| nano-banana-2 | PASS 13.1s | PASS 13.1s | 100% |
| nano-banana-pro | PASS 19.0s | FAIL 70s | 50% |
| flux-schnell | PASS 1.9s | PASS 2.4s | 100% |
| flux-dev-fp8 | PASS 3.9s | PASS 3.8s | 100% |
| flux-2-max | FAIL | PASS 30.1s | 50% |
| flux-2-pro | FAIL | PASS 18.4s | 50% |
| flux-2-klein-9b | FAIL | PASS 7.3s | 50% |
| flux-2-klein-4b | FAIL | PASS 5.9s | 50% |
The Stack
$ --driver nano-banana-2 — 13.1s, 100% reliable, highest quality
$ --draft flux-schnell — 2.2s, fast previews
$ --quality flux-dev-fp8 — 3.9s, speed-quality balance
$ --ip Flux.1 Dev + LoRA + ControlNet — character consistency
IP Pipeline
For strong character IP: train LoRAs on Flux.1 Dev with ControlNet-based workflows. Augment with Nano Banana 2 for precision passes or upscaling models for resolution. Watch cost at scale.
Train Flux.1 Dev LoRA
Control ControlNet
Augment NB2 / Upscale
Avoid BFL direct API