FLUX.2 LoRA API: finetune_id on BFL, not on Sume

On the BFL API you run a trained FLUX.2 LoRA by posting to a -finetuned endpoint with a finetune_id. Sume has no such path: its image docs say allowed_passthrough_parameters is empty for every endpoint in v1, so provider.options must be omitted or empty. Use reference images for consistency instead.

BFL details are from its LoRA inference page, which marks the feature as a public beta; Sume details are from Image generation, both read 2026-10-01.

How does BFL's LoRA inference work?

You train a LoRA against a FLUX.2 [klein] Base model, upload the .safetensors file in the BFL Dashboard, and the name you choose becomes your finetune_id. Each base model has a matching -finetuned endpoint, and the request adds finetune_id and finetune_strength to the base schema.

What happens if I send finetune_id to Sume?

The Sume docs state the rule: a request that sets a parameter the selected model does not list is rejected with 400 unsupported_parameter rather than silently dropped. finetune_id is not a listed parameter, so expect that error, and provider.options with provider keys is refused for the same reason.

What can I use for consistency instead?

Use image references. Sume's image models accept reference images where the model lists them; for example ChatGPT Image 2.5 supports up to 16 references and an optional mask_url. Read GET /v1/images/models to see what each model lists before you build around it. For a worked approach see consistent character AI image generator.

What should I do with a trained LoRA?

Keep serving it where it was uploaded. If your goal is a repeatable look rather than trained weights, move to references and a fixed prompt on Sume.