Reward-guided fine-tuning of Kiwi-Edit (5B)
Offline reward-ranked fine-tuning (RAFT): for each of 512 VEFX-Dataset training inputs Kiwi-Edit samples 8 candidates; Reward-LoRA trains on the
VEFX-Reward-4B top-1, Random-LoRA on a random candidate of the same pool (same data order, timesteps, noise). LoRA rank 64 on all DiT attention/FFN layers, 8 GPUs × batch 1,
Kiwi's own flow-matching objective. lr and checkpoint chosen on a 32-input dev split by an arm-agnostic rule; test = the 100 Best-of-K VEFX-Bench prompts × seeds 0–3.
← all results
Loading…
Selected examples on held-out VEFX-Bench prompts: the same source video, instruction and random seed, edited by the base Kiwi-Edit, by Random-LoRA (control: fine-tuned on randomly chosen samples) and by Reward-LoRA (fine-tuned on VEFX-Reward-selected samples). Scores: VEFX-Reward-32B (held-out evaluator; overall = GeoAgg, 1–4).
20 randomly sampled training inputs (of 512). Candidates ranked by VEFX-Reward-4B GeoAgg; Reward pick trains Reward-LoRA, Random pick trains Random-LoRA.