fix(api-proxy): offer a model hf-inference still serves

[rickstaa]

livepeer-example/flux-1-schnell is dead. Calling it returns HTTP 410 {"error":"The requested model is deprecated and no longer supported by provider hf-inference"}, so half of what the example advertises fails for anyone who runs it.

There is no same-modality replacement. Hugging Face's model index now lists exactly one text-to-image model on hf-inference (stabilityai/stable-diffusion-3-medium-diffusers, the one this example already uses). FLUX.1 schnell is still live on nscale, together, fal-ai and wavespeed, but each of those either puts the model in the request body (/v1/images/generations, so nginx can no longer pin it and the operator's token stops being spendable on exactly one model) or answers with a URL rather than the image, and they bill inference-provider credits rather than running on the free provider. Video is worse: hf-inference serves no text-to-video model at all, and every provider that does runs it as an async job, which a stock nginx cannot poll without becoming app code.

So the second capability stays in the image domain by reversing direction: google/vit-base-patch16-224 reads an image instead of making one. It keeps every property the example teaches — the model pinned in the URL, the caller's {"inputs": ...} body forwarded verbatim, hf-inference as the provider — and the price gap is real rather than decorative: one forward pass against a diffusion run, priced at 0.00001 against 0.0001. It also demonstrates itself, since the image SD3 paints is the image ViT reads back.

One wrinkle worth reviewing

call_runner accepts a JSON object or opaque bytes, and raises Live runner call expected JSON object, got list on anything else. Every Hugging Face inference task other than text-to-image answers with a bare JSON array, so this is not specific to ViT — it is what any config-only proxy hits in front of that API. A runner that is pure config can pin, inject, and relabel but cannot rewrite a body, so the route relabels the response Content-Type as text and the client parses it. The README says so in as many words, and the constraint is arguably worth teaching. If we would rather the SDK accepted a top-level array, that is a separate change and this route would drop the relabel.

Testing

Both paths, both capabilities:

• Offchain — SD3 wrote the JPEG, ViT answered llama at 0.9937 on that same file.
• On-chain (paid, arbitrum-one-mainnet) — both calls paid and accounted, fee=52781032608 for SD3 and fee=5278103260 for ViT, the 10x the config asks for. MAX_PRICE_PER_UNIT is unchanged and still clears the highest price.

Also updated: the root README transport column, the app README throughout, and the .env.example note that named flux as the cheaper capability.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.