flux2_klein_interleave_serve.yaml: run FastVideo as an InterleaveThinker-compatible generator service.
interleave_run.yaml: run the native planner -> generator -> critic loop through fastvideo interleave-run.
eval_prompts.jsonl: tiny prompt-set fixture for fastvideo interleave-eval.
interleave_single_prompt.py: run the native FastVideo interleave orchestrator with a fallback single-prompt planner and accept-all critic.

Compatibility Service

This service starts a FastVideo-backed image endpoint compatible with InterleaveThinker-style generator calls.

InterleaveThinker reward code posts JSON to /edit:

{
  "image": "<optional base64 png>",
  "prompt": "make the sky sunset orange",
  "num_inference_step": 4,
  "guidance_scale": 1.0,
  "width": 1024,
  "height": 1024
}

FastVideo returns:

{
  "success": true,
  "edited_image": "<base64 png>",
  "file_path": "/abs/path/to/output.png"
}

Run:

FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA \
  fastvideo interleave-serve --config examples/interleave/flux2_klein_interleave_serve.yaml

Then point InterleaveThinker at:

export EDIT_API_ENDPOINT=http://<host>:8011
export EDIT_MODEL_NAME=klein

The same app also exposes /generate and /v1/interleave/edit aliases.

Native Interleave Run CLI

Run the native orchestrator from a config file:

FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA \
  fastvideo interleave-run --config examples/interleave/interleave_run.yaml

Override common per-run fields from the CLI:

fastvideo interleave-run \
  --config examples/interleave/interleave_run.yaml \
  --prompt "create a clean product photo of a red ceramic mug" \
  --output-dir outputs/interleave_mug \
  --trace-path outputs/interleave_mug/trace.json

The example defaults to planner.kind: single_prompt and critic.kind: accept_all so it can smoke-test the FastVideo image backend without loading planner/critic checkpoints. To run the real InterleaveThinker loop, switch those blocks to interleave_thinker and set:

planner:
  kind: interleave_thinker
  init_from: InterleaveThinker/InterleaveThinker-Planner-8B
  processor_from: Qwen/Qwen3-VL-8B-Instruct
  trainable: false
  torch_dtype: bf16
  device_map: cuda:0
  attn_implementation: sdpa
  max_new_tokens: 2048

critic:
  kind: interleave_thinker
  init_from: InterleaveThinker/Critic-SFT-8B
  processor_from: Qwen/Qwen3-VL-8B-Instruct
  trainable: false
  torch_dtype: bf16
  device_map: cuda:0
  attn_implementation: sdpa
  max_new_tokens: 512

Loading the planner, critic, and image generator in one process can be memory intensive. On tighter GPUs, run the /edit compatibility service separately or load only the planner/critic in the native runner while the generator is served out-of-process.

Prompt-Set Evaluation

Run the same planner -> generator -> critic workflow over a JSONL prompt set:

FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA \
  fastvideo interleave-eval \
    --config examples/interleave/interleave_run.yaml \
    --prompts examples/interleave/eval_prompts.jsonl \
    --output-dir outputs/interleave_eval

Each prompt row writes outputs/interleave_eval/<id>/trace.json; the aggregate summary is written to outputs/interleave_eval/summary.json. The prompt-set loader accepts JSONL rows with id, prompt or instruction, optional initial_image_path, and arbitrary metadata fields. Use --resume to skip rows whose trace already exists and --limit N for small smoke runs.

Nano Banana / Gemini API Backends

The InterleaveThinker RL config can use closed-source Google models through lazy wrappers:

fastvideo.train.methods.rl.rewards.GeminiNanoBananaEditScorer generates edits with Nano Banana and scores them with Gemini.
fastvideo.entrypoints.interleave.generator.NanoBananaImageGeneratorBackend implements the same image backend protocol as the local FastVideo generator.

Install the optional SDK and provide a key only when using these API backends:

uv pip install -e ".[interleave-api]"
export GEMINI_API_KEY=...

Supported Nano Banana aliases are:

nano-banana -> gemini-2.5-flash-image
nano-banana-pro -> gemini-3-pro-image
nano-banana-2 -> gemini-3.1-flash-image

Native Single-Prompt Trace

Run:

FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA \
  python examples/interleave/interleave_single_prompt.py \
    --model-path black-forest-labs/FLUX.2-klein-4B \
    --prompt "a brushed steel espresso machine on a marble counter, morning window light" \
    --output-dir outputs/interleave_single_prompt

This writes an image plus trace.json. The trace records planner/generator/ critic attempts and omits base64 image payloads by default.