Eigen-Banana-Qwen-Image-Edit: Lightning-Fast Instruction-Based Image Editing with Pico-Banana-400K

TL;DR

Eigen-Banana-Qwen-Image-Edit is a LoRA fine-tune of Apache-2.0 Qwen-Image-Edit on Apple’s Pico-Banana-400K instruction-editing dataset, distributed as a plug-and-play adapter for Diffusers and DiffSynth-Studio. The release also ships a “Lightning” variant that generates edits in 4 inference steps, following the community Qwen-Image-Lightning recipe (FlowMatch-style scheduler + LoRA distillation). It is the first openly-released downstream model the wiki tracks that consumes Pico-Banana-400K’s SFT subset end-to-end and stacks a few-step distillation LoRA on top, on a different base than the Qwen-Image-Edit-2509 substrate the rest of the community has converged on.

Key claims

The release is a LoRA on Qwen-Image-Edit (not 2509) trained on Pico-Banana-400K, distributed for use in Diffusers’ QwenImage / Qwen-Image-Edit pipelines and in DiffSynth-Studio [§What this means for you].
A separate “Lightning” 4-step variant is produced via the community Qwen-Image-Lightning approach — FlowMatch-style scheduler with LoRA distillation — and exposes a 4-step preset for interactive / low-latency workloads [§Lightning variant].
Five qualitative comparison pairs are shown vs the base Qwen-Image-Edit (gallery bench composite, vintage film grade, race-bib text addition, indoor-event background replacement, smile retargeting) but no quantitative benchmark numbers are reported [§Comparison gallery].
The model is positioned within Eigen AI’s three-product stack: EigenTrain (the SFT / offline-RL / online-RL training platform used to produce this fine-tune), EigenInference (production-grade quantization / sparsity / distillation / routing for serving), and EigenDeploy (autoscaling deployment) [§Platform].
The hosted demo is claimed to be “much faster without any quality loss” than the open-source LoRA branch, implying additional non-released inference optimizations sit between the LoRA and the hosted endpoint [§Hosted demo].

Method

The pipeline is a two-LoRA stack on a frozen Qwen-Image-Edit backbone. Stage 1 is an SFT LoRA: Qwen-Image-Edit is fine-tuned on the success subset of Pico-Banana-400K (Apple’s ~258K (image, instruction, edited-image) triplets distilled from Nano-Banana over OpenImages photos with a Gemini-2.5-Pro four-criterion judge at threshold ~0.7). Stage 2 is a Lightning distillation LoRA following the community Qwen-Image-Lightning recipe, which uses a FlowMatch-style scheduler and LoRA distillation to compress the model to 4 inference steps. The two LoRAs are intended to be loaded together at inference. The hosted demo additionally applies unreleased EigenInference optimizations (quantization / sparsity / distillation / routing per the platform pitch); these are not characterized in the post. No training-data filtering, evaluation, ablation, or per-edit-type success-rate analysis is reported beyond Apple’s original Pico-Banana-400K coverage.

Results

No benchmark numbers are reported. The post is qualitative: five before/after image pairs (gallery bench composite, vintage film grade, bib-text addition, indoor-event background replacement, subtle smile) are shown to demonstrate that the 4-step Lightning variant preserves the base Qwen-Image-Edit’s edit fidelity on edits the original model handles well. Reported deliverables: open-source LoRA weights (Diffusers + DiffSynth-Studio compatible), a hosted faster-inference demo, and a sales path to the underlying EigenTrain / EigenInference / EigenDeploy stack.

Why it’s interesting

This is the first openly-released downstream model the wiki tracks that consumes Pico-Banana-400K — Apple’s recipe of “Nano-Banana as teacher + Gemini-2.5-Pro as judge + OpenImages as source” (Pico-Banana-400K: A Large-Scale Dataset for Text-Guided Image Editing) — to fine-tune an Apache-2.0 base. That makes it a useful data point on whether the Pico-Banana-400K SFT subset is enough by itself to lift a strong open-source editor, or whether the preference-pair subset (which would need a downstream DPO / reward-model stage) is doing most of the actual quality work in the recipe. The stacking pattern — a task-specific LoRA + a Lightning few-step distillation LoRA on a frozen image-edit DiT — is the same recipe the Qwen-Image-Edit-2509 community has converged on for unrelated tasks like camera control (Qwen-Edit-2509-Multiple-angles LoRA — community camera-control adapter for Qwen-Image-Edit-2509) and photo restoration (Qwen-Edit-2509-Upscale-LoRA — community photo restoration adapter for Qwen-Image-Edit-2509), now extended to general instruction editing on the prior 2508-era base. The Lightning component itself sits in the same lineage as the from-scratch and DMD-style few-step distillation work cataloged in Diffusion Distillation (ArcFlow, DP-DMD, TDM), but with no published technical details beyond “community Qwen-Image-Lightning approach” — so it is a deployment data point, not a methodological one.